Efficient Two-Step Estimation via Targeting David T. Frazier and Eric Renault

advertisement
Efficient Two-Step Estimation via Targeting
David T. Frazier⇤and Eric Renault†
April 24, 2015
Abstract
The standard description of two-step extremum estimation amounts to plugging in a
first step estimator of nuisance parameters in order to simplify the optimization problem
and then to deduce a user friendly estimator for the parameters of interest. This two-step
procedure often induces an efficiency loss with respect to estimation of the parameters of
interest. In this paper, we consider a more general setting where we do not necessarily
have such thing as nuisance parameters but rather awkward occurrences of the parameters of interest. By awkward, we mean that within the estimating equations for a vector
of unknown parameters of interest ✓ , some occurrences of ✓, encapsulated by a vector
⌫(✓), may be computationally tricky. Then, it is still the case that prior knowledge of the
unknown auxiliary parameters ⌫ = ⌫(✓) would make inference on ✓ much simpler, and it
is this fact that motivates the two-step approach developed in this paper. The efficiency
problem is more difficult than for the case of standard nuisance parameters since even
the (infeasible) approach of plugging in the true unknown value of ⌫ = ⌫(✓) may not
allow efficiency, since it overlooks the information about ✓ contained in the awkward occurrences ⌫(✓). Moreover, we stress that standard ways to restore efficiency for two-step
procedures may not work due to a consistency issue; when setting the focus on a first step
estimator for only some of the occurrences ⌫ = ⌫(✓) of the unknown parameters ✓, global
identification may be lost. To alleviate this issue, we develop a targeting strategy that
enforces consistency and achieves efficiency. Such difficult occurrences ⌫(✓) of the parameters, which are a nuisance when it comes to solving estimating equations, are present in
many financial econometrics applications, often handled by indirect inference. Leading
examples are asset pricing models with latent variables (their observation would make
estimation much simpler), models where it is simpler to first set the focus of inference
on marginal distributions (multivariate GARCH, copulas), models with highly nonlinear
objective functions, etc. Based on targeting and penalization of the auxiliary parameters,
we propose a new two-step estimation procedure that leads to stable and user-friendly
computations. Moreover, estimators delivered in the second step of the estimation procedure are asymptotically efficient. We compare this new method with existing iterative
methods in the framework of copula models and asset pricing models. Simulation results
illustrate that this new method performs better than existing iterative procedures and is
(nearly) computationally equivalent.
⇤
†
Department of Econometrics and Business Statistics, Monash University. email: david.frazier@monash.edu
Department of Economics, Brown University. email: eric renault@brown.edu
1
Keywords : Targeting, Penalization, Multivariate Time Series Models, Asset Pricing, Implied
States.
2
1
Introduction
The standard treatment of two-stage estimation (see e.g. Pagan, 1986 or Newey and McFadden,
1994, section 6) is generally motivated by the following sequence of arguments as coined by
Pagan (1986):
(i) Econometricians are often faced with the troublesome problem that “in order to estimate
the parameters they are ultimately interested in, it becomes necessary to quantify a number
of nuisance parameters (...) it is the presence of these parameters which converts a relatively
simple computational problem into a very complex one”.
(ii) ”Because estimation would generally be easy if the nuisance parameter were known, a
very common strategy for dealing with them has emerged: they are replaced by a nominated
value which is estimated from the data”. Then, the key issue for asymptotic theory is to assess
the e↵ect of first-step estimators on second-step standard errors (see Newey and McFadden,
1994, subsection 6.2) and the most favorable situation is when ignoring the first step would be
valid: the asymptotic distribution on the second-step estimator for the parameters of interest
does not depend on the first step estimator for the nuisance parameters and would have been
the same whether the nuisance parameters had been known upfront.
Our focus of interest in this paper is germane to the above one but more general. The
main di↵erence is that we do not necessarily have such thing as nuisance parameters but rather
awkward occurrences of the parameters of interest. By awkward, we mean that within the
estimating equations for a vector of unknown parameters of interest ✓, some occurrences of ✓
may be computationally tricky, either due to the complexity of the relationship, or numerical
instability, or both. In order to disentangle these unpleasant occurrences from user-friendly
ones, we denote the sample-based estimating functions as qT [✓, ⌫(✓)], where ⌫(✓) encapsulates
all the occurrences of ✓ considered as somewhat awkward while T stands for the sample size.
Generally speaking, our estimator of interest is ✓ˆT defined as a zero of the vector function
fT (✓) = qT [✓, ⌫(✓)].
Note that, this general framework obviously encompasses the standard nuisance parameter
setting described above. If, within the vector ✓ of unknown parameters, we distinguish some
parameters of interest, denoted by ✓1 , and some nuisance parameters, denoted by ✓2 , such that
✓ = (✓10 , ✓20 )0 and ⌫(✓) = ✓2 , we are back to the standard case as far as efficient estimation of
✓1 is concerned. Note that, up to a slight change of notation, our setup nests the case where
the function ⌫(✓) would be a sample dependent one ⌫T (✓), for instance because ⌫(✓) shows up
after some nuisance parameters have been profiled out. Up to a specific discussion on how to
accommodate this case (see the Appendix), the simpler notation ⌫(✓) will be kept throughout.
Our leading example will be the case of an extremum estimator
✓ˆT = arg max QT [✓, ⌫(✓)],
✓
(1)
so that the estimating equations correspond to first order conditions:
qT [✓, ⌫(✓)] =
@QT [✓, ⌫(✓)] @⌫ 0 (✓) @QT [✓, ⌫(✓)]
+
.
@✓
@✓
@⌫
(2)
✓ˆT may be the MLE if the function QT [✓, ⌫(✓)] is a well-specified (log)likelihood function. More
3
generally, we will see ✓ˆT throughout as our benchmark estimator for the purpose of asymptotic
efficiency.
We highlight two important classes of examples in this paper. First, in Section 4, we
consider a class of additively separable log-likelihood functions that are usually encountered
in the so-called“estimation from likelihood of margins” (see e.g. Joe, 1997). In this setting,
the unknown parameters, components of ✓, can be split into two parts ✓ = (✓10 , ✓20 )0 , where
✓1 characterizes the likelihood of the margins and ✓2 characterizes the dependence between
components, let’s say the “cross-dependence”, through some link functions (typically linear
correlations or copulas). However, the link function describing the cross-dependence applies
to data components that have been first standardized using the knowledge of ✓1 . In other
words, the part of the likelihood capturing cross-dependence also involves the parameters ✓1
that describe the marginal distributions. Such occurrences of ✓1 are an example of the awkward
occurrences mentioned earlier, in that these situations can be difficult to deal with in practice;
i.e., ⌫(✓) = ✓1 corresponds to the occurrences of ✓1 in the cross-dependence portion of the
log-likelihood. Fortunately, a consistent user friendly estimator of ✓1 is available from the
likelihood of the margins and can be plugged into the cross-dependence portion in order to
estimate ✓2 . This approach is actually popular for the estimation of nonlinear multivariate
time series models like multivariate GARCH or copulas models. However, as explained below,
the simplicity obviously entails an efficiency loss since the information in the cross-dependence
model about the margin parameters ✓1 has been overlooked.
In Section 5, we consider nonlinear models in which observable variables are viewed as
functions of some latent variables. Typically, the latent model, which is characterized by a
vector of unknown parameters ✓, specifies a Markov process for the state variables and defines
their (possibly nonlinear) transition equation. Such an approach becomes difficult when the
measurement equation of this non-linear state space model, which is the function that relates
observable variables to latent ones, also depends on the same unknown parameters through
a vector ⌫(✓). While it would have been relatively easy to estimate ✓ from the observations
on the latent variables, inference using available observations is complicated by the additional
awkward occurrence of ✓, namely ⌫(✓), in the transformation from latent to observable variables. It is worth noting that the issue we have in mind is not really about filtering latent
variables because we actually consider a case where the relationship latent-observable is oneto-one. Hence, backing out the latent variables from observations would have been easy if not
polluted by the additional occurrence of unknown parameters in the measurement equation.
This kind of situation is common in modern arbitrage-based asset pricing models with hedging of various sources of risk defined by an underlying model of state variables. Latent state
variables are common factors, the dynamics of which characterizes the dynamics of observed
yields or derivative asset prices. Since the measurement equation, typically an arbitrage-based
asset pricing formula, is one-to-one, we can, following Pan (2002), dub ”Implied States” the
value of latent variables that can be backed out from observations for a given value of parameters ⌫(✓). From this general intuition, Pan (2002) has extended the approach put forward by
Renault and Touzi (1996) (and later revisited by Pastorello, Patilea and Renault (2003)) to
devise the so-called “Implied States GMM” estimator. Again, the simplicity of this strategy
also comes at the cost of an efficiency loss since the information content about ✓ brought by its
awkward occurrence ⌫(✓) is overlooked in this procedure. As Pan (2002) put it, “the efficiency
4
of this “optimal instrument”scheme is limited in that (...) we sacrifice efficiency by ignoring
the dependence of (t) on ✓, ” spot volatility (t) backed out from option price being for her
the implied state.
We are then generally faced with the following trade o↵ between asymptotic efficiency and
computational cost (both in terms of computational complexity and stability). On the one
hand, we still contemplate that estimation would be easy if the awkward part ⌫(✓) were known.
Therefore, there is still some rationale to estimate it in a first stage, that is, if ✓0 stands for the
true unknown value of ✓, to replace ⌫(✓0 ) by a consistent sample counterpart ⌫˜T .
On the other hand, it is well known (see Newey and McFadden, 1994 for a discussion)
that the two-step estimator obtained by plugging in the first-step consistent estimator ⌫˜T of
the nuisance parameters would be inefficient in general. However, we want to stress that in
our more general case where ⌫ is not necessarily a nuisance parameter but may be a known
function ⌫(✓) of parameters of interest, there is even no reason to believe that we would get a
more accurate estimator by computing the infeasible estimator ✓˘T , the solution of
qT [✓˘T , ⌫(✓0 )] = 0.
(3)
On the contrary, there are many circumstances (see Pastorello et al., 2003 and references
therein) in which the infeasible estimator ✓˘T is actually less accurate than ✓ˆT . The efficiency
loss is due to the fact that the computation of ✓˘T disregards the information about ✓ contained
in the function ⌫(✓) (see also Crepon et al., 1997 for a similar remark in a GMM context).
More precisely, the efficient estimator ✓ˆT is asymptotically equivalent to:

1
@qT 0
@qT 0
0
0 @⌫
0
[✓ , ⌫(✓ )] +
[✓ , ⌫(✓ )] 0 (✓ )
qT [✓0 , ⌫(✓0 )],
0
0
@✓
@⌫
✓
while the infeasible estimator ✓˘T is asymptotically equivalent to:

1
@qT 0
0
[✓
,
⌫(✓
)]
qT [✓0 , ⌫(✓0 )].
@✓0
Two standard strategies are available in the literature to address this efficiency issue. A
first possibility, as recently developed by Fan, Pastorello and Renault (2015) (hereafter, FPR)
(k)
is to devise a sequence of estimators ✓ˆT , k = 1, 2... from a feasible counterpart of (3)
qT [✓ˆT
(k+1)
, ⌫(✓ˆT )] = 0,
(k)
(4)
with, for instance, the aforementioned consistent first-step estimator ✓˜T as the initial value
(1)
(✓ˆT = ✓˜T ). Following a seminal paper by Song, Fan and Kalbfleich (2005) (hereafter, SFK)
who had proposed a simplified version of this strategy in the particular case of separable loglikelihood functions (see Section 4 below), with the algorithm (4) being dubbed “Maximization
by Parts” (MBP hereafter). The nice thing with (4) is that each step of the iteration to compute
(k+1)
(k)
✓ˆT
from ✓ˆT is no more computationally demanding than the solution of (3). Moreover, by
contrast with (3), this iterative procedure may allow us to reach efficiency since, when the
(1)
iterative procedure (4) has a limit ✓ˆT , this limit must coincide with the efficient estimator
5
✓ˆT . However, it is worth realizing that the required contraction mapping property to secure
convergence of (4) is not in general fulfilled in finite samples. Therefore, a feasible efficient
estimator relies upon the choice of a tuning parameter k(T ), going to infinity at a sufficient
(k(T ))
rate with the sample size T , in order to obtain an estimator ✓ˆT
that is asymptotically
ˆ
equivalent to ✓T . This may obviously come with the computational cost of a large number k(T )
of iterations, especially when the required population contraction mapping property is hardly
fulfilled. Needless to say, the situation is even worse when it is not fulfilled at all, as illustrated
in Section 4 below.
The main goal of this paper is to promote a new efficient two-step procedure that does not
require the contraction mapping property. We will argue that even though its second-step may
be more computationally involved than each step of MBP, it keeps some of its simplicity, in
particular by comparison with the brute force computation of the efficient estimator ✓ˆT . Our
efficient two-step procedure is actually an extension of a two-step extremum estimator first
proposed by Trognon and Gourieroux (1990). The key intuition is to correct the naive twostep objective function QT [✓, ⌫˜T ] to compensate for the inefficiency caused by plugging in the
first-step consistent estimator ⌫˜T . Our proposed extremum estimator would then be
✓ˆText = arg max Q̃T [✓, ⌫˜T ],
(5)
✓
with
Q̃T [✓, ⌫˜T ] = QT [✓, ⌫˜T ] +
@QT [✓, ⌫˜T ]
. [⌫(✓)
@⌫ 0
and

P lim JT (✓0 ) +
T =1
⌫˜T ]
1
[⌫(✓)
2
⌫˜T ]0 JT (✓) [⌫(✓)
⌫˜T ]
(6)
@ 2 QT [✓0 , ⌫(✓0 )]
= 0.
@⌫@⌫ 0
We show that, when consistent, the estimator ✓ˆText is asymptotically equivalent to the efficient
estimator ✓ˆT . The main intuition for this result is that, up to the occurrence of unknown ✓ inside
the matrix JT (✓), the first order conditions of the maximization program (5) can be seen as a
linearization of first order conditions (2) of the efficient program (1), namely, linearization with
respect to ⌫ in the neighborhood of the first-step estimator ⌫˜T . Then, the efficiency argument
will be based on a generalization of an argument extensively studied by Robinson (1988). In
this seminal paper, general efficiency comparisons are led between roots of rival estimating
equations, in particular, as provided by local linearizations. However, we point out a difficulty
that seems to have been overlooked in the literature so far. When linearization around a
preliminary consistent estimator is applied to a vector of estimating equations, like, fT (✓) =
qT [✓, ⌫(✓)], but linearization is performed only with respect to the second set of occurrences
of ✓ (the so-called awkward occurrences within ⌫(✓)), the fact that fT (✓) may also depend
nonlinearly on ✓ through first occurrences, say ✓ = ✓⇤ in qT [✓⇤ , ⌫(✓)], can impair the consistency
of the estimator defined as the root of this (partially) linearized estimating equation. More
precisely, local identification is granted but not global identification.
Our proposed hedge against this risk is the addition of a penalty term ↵T k⌫(✓) ⌫˜T k2 to
the (partially linearized) estimating equations, with a tuning parameter ↵T going to infinity
6
slower than the rate of convergence of our initial estimator ⌫˜T . In other words, both the MBP
approach and our new two-step procedure with penalized partial linearization come with the
cost of a tuning parameter. While the MBP approach requires choosing the number k(T ) of
iterations, instead we will have to choose the rate of divergence ↵(T ) of the penalty weight. We
will see, for instance, that in the standard case where all estimators are root-T consistent, a rate
T 1/4 is well suited. Moreover, we will propose two di↵erent two-step procedures, both based
on partial linearization, depending upon whether we have a first-step consistent estimator ⌫˜T ,
only of the unpleasant parameters, or we have at our disposal an initial consistent estimator
✓˜T for the whole parameter vector ✓ (and then ⌫˜T = ⌫(✓˜T ) ). Of course, in the latter case, the
2
penalty term could instead be ↵T ✓ ✓˜T and should be able to enforce global identification
in even more general circumstances.
The paper is organized as follows. The proposed extension of the Trognon and Gourieroux
(1990) efficient two-step procedure is studied in Section 2. Our general result explains why
some well known two-step estimators are efficient, in spite of the appearance to the contrary:
Hatanaka (1974) for a dynamic regression model, Gourieroux, Monfort and Renault (1996)
for a GMM estimator. It is worth stressing that efficiency is warranted in these two specific
examples because consistency is not an issue. However, we also point out other examples, such
as, nonlinear least squares and GMM, where consistency is not warranted, except if one uses
the penalty strategy that we have devised through first order conditions. Robinson’s (1988)
comparison of estimators is developed in Section 3. It allows us to propose two di↵erent penalized two-step estimators, depending on whether one has at her disposal a first-step consistent
estimator of ✓0 or only of ⌫(✓0 ). Section 4 sets the focus on the separable estimation problem
with a detailed comparison with MBP, both analytically and through Monte Carlo experiments
in the framework of a copula example. Section 5 addresses the general implied states issue,
both in the context of maximum likelihood and GMM as well. Again, we are able to provide
a detailed comparison with MBP, both analytically and through Monte Carlo experiments, in
the simple framework of Merton’s credit risk model. Concluding remarks are given in Section
6.
Mathematical proofs, regularity conditions and detailed Monte Carlo evidence are all gathered in the Appendix.
2
2.1
An efficient two-step extremum estimator
General framework
Let ⇥ ⇢ Rp be a compact parameter space, and ✓0 the true unknown value of ✓. Additional
parameters ⌫ are defined by some continuous function ⌫(.) from ⇥ to some subset of Rq .
We assume that the extremum estimator ✓ˆT of ✓, defined by (1), is a consistent asymptotically
normal estimator of ✓0 . In addition, we assume the following standard regularity conditions
are satisfied.
Assumption A1: There is a real-valued deterministic function Q1 [., .], continuous on ⇥ ⇥
7
and such that:
(i) PlimT =1 sup |Q1 [✓, ⌫(✓)]
QT [✓, ⌫(✓)]| = 0 and
✓2⇥
(ii) ✓0 = arg max Q1 [✓, ⌫(✓)].
✓2⇥
Assumption A2: The following are satisfied
(i) ⌫(.) is twice continuously di↵erentiable on the interior of ⇥.
(ii) ✓0 2 Int(⇥), interior set of ⇥, and ⌫ 0 = ⌫(✓0 ) 2 Int( ), interior set of .
˚ ˚ and with qT [✓, ⌫(✓)]
(iii) The function QT [✓, ⌫] is twice continuously di↵erentiable on ⇥⇥
defined by (2)
p
(1) T qT [✓0 , ⌫(✓0 )] !d @[0, I0 ]
n
o
0
0 )]
0
0 )] @⌫(✓ 0 )
(2) PlimT =1 @qT [✓@✓,⌫(✓
+ @qT [✓@⌫,⌫(✓
. @✓0
= H0
0
0
In addition, we maintain the following high-level assumptions.
Assumption A3: (✓ˆT0 , ⌫˜T0 )0 is a root-T consistent asymptotically normal estimator of (✓00 , ⌫ 00 )0 and
PlimT =1 sup JT (✓) +
✓2⇥
@ 2 QT [✓, ⌫˜T )]
= 0.
@⌫@⌫ 0
The focus of interest in this section is the comparison of the efficient estimator ✓ˆT , with the
two-step alternative ✓ˆText defined in the introduction. For the sake of interpretation, it is worth
comparing ✓ˆT and ✓ˆText with the infeasible estimator
✓ˆT⇤ = arg max Q̃0T [✓, ⌫˜T ],
✓2⇥
where
Q̃0T [✓, ⌫˜T ] = QT [✓, ⌫˜T ] +
@QT [✓, ⌫˜T ]
. [⌫(✓)
@⌫ 0
⌫˜T ]
1
[⌫(✓)
2
⌫˜T ]0 JT (✓0 ) [⌫(✓)
⌫˜T ] .
We are then able to prove the following result.
Theorem 2.1: Under the maintained assumption that they are all root-T consistent, the three
estimators ✓ˆT , ✓ˆText and ✓ˆT⇤ are asymptotically equivalent.
We stress that, as announced in the introduction, it is only the careful analysis of the
first order conditions (see section 3 below) that will allow us to devise a proper penalty to
ensure consistency of our two-step estimators. It is, however, worth interpreting them further
to discern the reason why the two-step approach is not responsible for any efficiency loss. We
8
can do that at least in the only case considered by Trognon and Gourieroux (1990), namely the
case of a genuine nuisance parameter
✓ = (✓10 , ✓20 )0 , ⌫(✓) = ✓2 .
Then, the modified objective function becomes
Q̃T [✓, ⌫˜T ] = QT [✓, ⌫˜T ] +
@QT [✓, ⌫˜T ]
. [✓2
@⌫ 0
⌫˜T ]
1
[✓2
2
⌫˜T ]0 JT (✓) [✓2
⌫˜T ] ,
and the parameters of interest for efficient estimation are included in the sub-vector ✓1. With
this point in mind, we can set the focus on an even simpler two-step estimator obtained as the
maximizer of the following simplified objective function, where for sake of avoiding confusion
about partial derivatives, we use two di↵erent notations for the same first-step estimator
✓˜2,T = ⌫˜T
@QT [✓1 , ✓˜2,T , ⌫˜T ]
Q̆T [✓, ⌫˜T ] = QT [✓1 , ✓˜2,T , ⌫˜T ] +
. [✓2
@⌫ 0
1
[✓2
2
⌫˜T ]
⌫˜T ]0 JT (✓1, ✓˜2,T ) [✓2
⌫˜T ]
Then, it is easy to profile ✓2 out of Q̆T [✓, ⌫˜T ]
h
i
@ Q̆T [✓, ⌫˜T ]
= 0 , ✓2 = ⌫˜T + JT (✓1, ✓˜2,T )
@✓2
1
@QT [✓1 , ✓˜2,T , ⌫˜T ]
.
@⌫
Plugging the above value of ✓2 into Q̆T [✓, ⌫˜T ], we can concentrate the objective function with
respect to the nuisance parameters ⌫(✓) = ✓2 and obtain the following profile objective function
i
1 @QT [✓1 , ✓˜2,T , ⌫˜T ] h
˜
˜
Q̆c,T [✓1 , ⌫˜T ] = QT [✓1 , ✓2,T , ⌫˜T ] +
JT (✓1, ✓2,T )
2
@⌫ 0
1
@QT [✓1 , ✓˜2,T , ⌫˜T ]
.
@⌫
For sake of interpretation, let us consider instead the infeasible objective function and its
profile counterpart. Then, the concentrated score vector is
i
@ Q̆0c,T [✓1 , ⌫˜T ]
@QT [✓1 , ✓˜2,T , ⌫˜T ] @ 2 QT [✓1 , ✓˜2,T , ⌫˜T ] h
0 ˜
=
+
J
(✓
,
✓
)
T 1 2,T
@✓1
@✓1
@✓1 @⌫ 0
1
@QT [✓1 , ✓˜2,T , ⌫˜T ]
.
@⌫
From the definition of the matrix JT (✓0 ), we can then deduce that
PlimT =1
@ Q̆0c,T [✓10 , ⌫ 0 ]
= 0.
@✓1 @⌫ 0
(7)
Equation (7) is precisely the standard condition (see e.g Newey and McFadden, 1994, formula (6.6) p 2179) to ensure that the asymptotic distribution of the estimator of the parameters
of interest ✓1 does not depend on the asymptotic distribution of the estimator for the nuisance
parameters ⌫. This provides clear intuition as to why Theorem 2.1. works, at least in the particular case considered by Trognon and Gourieroux (1990): the modification of the objective
9
function in (6) has been devised precisely to restore the asymptotic independence between the
two kinds of parameters. However, the main contribution of the paper is to provide a much
more general setup for efficient two-step estimation through the use of penalized estimating
equations. This penalty amounts to a slight twist (via targeting) on the two-step estimator
✓ˆText , in order to ensure its consistency. Then, its asymptotic equivalence with ✓ˆT , as stated in
Theorem 2.1., ensures its asymptotic efficiency. As far as the equivalence between ✓ˆText and ✓ˆT⇤
is concerned, note that this is germane to the equivalence between two-step efficient GMM and
continuously updated GMM, as first put forward by Hansen et al. (1996).
2.2
Application to nonlinear regression
In this subsection, we consider the example of nonlinear least squares. Note that while we
consider only ordinary least squares, weighted least squares would not introduce any specific
difficulty. Joint estimation of models for conditional mean and variance using Gaussian QMLE
(Bollerslev and Wooldridge, 1992) would also fit in this class of examples. Thus, for sake of
notational simplicity, let us just consider the following objective function
QT [✓, ⌫(✓)] =
T
1X
[yt
T t=1
g(xt , ✓, ⌫(✓))]2 ,
where g(., ., .) is a known function such that
g(xt , ✓0 , ⌫(✓0 )) = E[yt |xt ] .
(8)
Hence, the maintained identification assumption is
E[yt
g(xt , ✓, ⌫(✓)) |xt ] = 0 , ✓ = ✓0 .
(9)
Then,
T
@QT [✓, ⌫(✓)]
2 X @g(xt , ✓, ⌫(✓))
=
[yt
@⌫
T t=1
@⌫
@ 2 QT [✓, ⌫(✓)]
=
@⌫@⌫ 0
g(xt , ✓, ⌫(✓)] ,
T
2 X @g(xt , ✓, ⌫(✓)) @g(xt , ✓, ⌫(✓))
.
T t=1
@⌫
@⌫ 0
T
2 X @ 2 g(xt , ✓, ⌫(✓))
+
[yt
T t=1
@⌫@⌫ 0
g(xt , ✓, ⌫(✓)] .
However, by applying (8), we can choose the following consistent estimator for the Hessian
matrix with respect to the parameters ⌫
JT (✓) =
T
2 X @g(xt , ✓, ⌫˜T ) @g(xt , ✓, ⌫˜T )
.
.
T t=1
@⌫
@⌫ 0
10
With this choice, the modified extremum estimator is obtained as the maximizer of
Q̃T [✓, ⌫˜T ] = QT [✓, ⌫˜T ] +
T 
1X
=
yt
T t=1
@QT [✓, ⌫˜T ]
1
. [⌫(✓) ⌫˜T ]
[⌫(✓) ⌫˜T ]0 JT (✓) [⌫(✓)
@⌫ 0
2
2
@g(xt , ✓, ⌫˜T )
g(xt , ✓, ⌫˜T )
.
[⌫(✓)
⌫
˜
]
.
T
@⌫ 0
⌫˜T ] (10)
In other words, while the estimator defined as the solution to
min
✓
T
X
[yt
g(xt , ✓, ⌫˜T )]2
t=1
is not efficient in general, we can restore efficiency by the additional term in (10). The fact that
a nonlinear regression model can be efficiently estimated after linearization of the regression
function around a first-step consistent estimator has been known since Hartley (1961). However,
it must be kept in mind that the case (10) is more general because we consider only a partial
linearization so as to deal with the nasty occurrences ⌫(✓) in g(·). As a result, efficiency is
warranted only when consistency is enforced, which may take the penalty strategy developed
in Section 3. To see this, note that the identification assumption (9) does not say that
E[yt |xt ] = g(xt , ✓, ⌫(✓0 ))
@g(xt , ✓, ⌫(✓0 )) ⇥
. ⌫(✓)
@⌫ 0
⇤
⌫(✓0 ) ) ✓ = ✓0 .
The role of targeting will be to enforce the equality ⌫(✓) = ⌫(✓0 ) so that the implication
above becomes a consequence of the identification assumption (9). Fortunately, there are cases
where penalty/targeting is not needed because consistency is directly implied. Trognon and
Gourieroux (1990) point out the example of Hatanaka’s (1974) two-step estimator for a dynamic
adjustment model with autoregressive errors. With obvious notations, the model is
y t = ↵ 1 y t 1 + ↵ 2 z t + ut
ut = ut 1 + "t
and is generally rewritten as
yt
yt
1
= ↵1 (yt
1
yt 2 ) + ↵2 (zt
z t 1 ) + "t .
Thus, we end up with a nonlinear regression model that can be rewritten in the notational
system of (8)
yt = g (xt , ↵1 , ↵2 , ⌫(✓)) + "t
xt = (zt , yt 1 ), ✓ = (↵1 , ↵2 , )0 , ⌫(✓) = .
However, a key remark is that the regression function, albeit nonlinear, is linear with respect to
⌫ when the friendly occurrence of ✓ is fixed. Therefore, this partial linearization with respect
to ⌫ does not cause consistency to break down. Theorem 2.1 can be directly applied to confirm
11
that Hatanaka’s (1974) two-step estimator is efficient.
2.3
Application to GMM
We now contemplate the case of a parameter identified through H moment restrictions with
two kinds of occurrences for the parameters:
E['t (✓, ⌫(✓))] = 0 , ✓ = ✓0 .
(11)
Moment restrictions of the form in (11), and their possible applications, are discussed in more
details in Section 5 within the setting of implied states GMM. In this section, however, we give
a general discussion in the context of modified two-step extremum estimators.
When working with (11), we typically have in mind estimators defined from the criterion
function
QT [✓, ⌫(✓)] = '¯T (✓, ⌫(✓))0 WT '¯T (✓, ⌫(✓))
where
T
1X
'¯T (✓, ⌫(✓)) =
't (✓, ⌫(✓))
T t=1
and WT is some positive definite sequence of matrices. Note that, in order to obtain an estimator ✓ˆT , defined by (1), that reaches the semiparametric efficiency bound, the sequence
WT should provide
a consistent
hp
i estimator for the inverse of the long term variance matrix
0
0
limT =1 V ar T '¯T (✓ , ⌫(✓ )) . However, this issue is irrelevant for us as we only discuss how
to obtain estimators that are asymptotically equivalent to ✓ˆT , irrespective of its efficiency.
From the definition of QT [✓, ⌫(✓)],
@QT [✓, ⌫(✓)]
=
@⌫
@ 2 QT [✓, ⌫(✓)]
=
@⌫@⌫ 0
@ '¯T (✓, ⌫(✓))0
WT '¯T (✓, ⌫(✓))
@⌫
@ '¯T (✓, ⌫(✓))0
@ '¯T (✓, ⌫(✓))
2
WT
@⌫
@⌫
H
2
X @ '¯h;T (✓, ⌫(✓))
2
.Wh.;T '¯T (✓, ⌫(✓))
0
@⌫@⌫
h=1
2
where Wh.;T stands for the hth row of WT . Then, we can choose the following consistent
estimator for the Hessian matrix with respect to the parameters ⌫
JT (✓) = 2
@ '¯T (✓, ⌫(✓))0
@ '¯T (✓, ⌫(✓))
WT
.
@⌫
@⌫ 0
12
With this choice, the modified extremum estimator is obtained as the maximizer of
@QT [✓, ⌫˜T ]
1
Q̃T [✓, ⌫˜T ] = QT [✓, ⌫˜T ] +
. [⌫(✓) ⌫˜T ]
[⌫(✓) ⌫˜T ]0 JT (✓) [⌫(✓) ⌫˜T ]
0
@⌫
2


0
@ '¯T (✓, ⌫˜T )
@ '¯T (✓, ⌫˜T )
=
'¯T (✓, ⌫˜T ) +
. [⌫(✓) ⌫˜T ] WT '¯T (✓, ⌫˜T ) +
. [⌫(✓)
0
@⌫
@⌫ 0
(12)
⌫˜T ]
In other words, while the solution of
min ['¯T (✓, ⌫˜T )]0 WT ['¯T (✓, ⌫˜T )]
(13)
✓
would not be equivalent to ✓ˆT in general, we can restore equivalence (and efficiency in the sense
of ✓ˆT ) by using the additional term in (12). However, since (similarly to the former subsection)
it is only a partial linearization of the moment conditions, consistency may not be warranted.
To see why consistency may be an issue, note that the identification assumption (11) does
not say that

⇤
@'t (✓, ⌫(✓0 )) ⇥
E 't (✓, ⌫(✓0 )) +
. ⌫(✓) ⌫(✓0 ) = 0 =) ✓ = ✓0 .
(14)
0
@⌫
The role of targeting in this context is to enforce the equality ⌫(✓) = ⌫(✓0 ) so that the implication in (14) becomes a consequence of the identification assumption (11).
Fortunately, there are cases where the penalty/targeting is not needed because consistency
is directly implied. Gourieroux et al. (1996) consider the case where the vector of moment
conditions can be split in two parts, with only the second one depending on ⌫:
0
't (✓, ⌫(✓)) = ['1t (✓)0 , '2t (✓, ⌫(✓))0 ] .
(15)
Then, the implication (14) is obviously warranted when the first set of moment conditions is
sufficient to identify ✓, which is typically the case considered by Gourieroux et al. (1996). In
this case, Theorem 2.1. ensures efficiency of the modified two-step estimator.
Interestingly enough, the efficient two-step estimator proposed by Gourieroux et al. (1996)
may be di↵erent from ✓ˆText . It is only when the first set of moment conditions '1t (✓) is linear
with respect to ✓ that they will numerically coincide (see section 2.6 in Gourieroux et al., 1996).
In the general case, their two-step efficient estimator is not based on a (partial) linearization
but on minimizing the norm of the moment vector '¯T (✓, ⌫˜T ), where the weighted matrix is a
suitably twisted version of the estimator for the inverse of the long term variance matrix. Note
that, we know from (13) that efficiency cannot be met without such a twist. Moreover, our
equivalence result is more general since not only does it apply to general moment conditions
(not only in the form (15)) but it does not assume that WT is a consistent estimator of the
inverse of the long term variance matrix.
13
3
Stochastic di↵erences for linearized estimating equations
We first state our general result concerning roots of linearized estimating equations, which
extends Theorem 2 of Robinson (1988). Then, in a second subsection, we provide two more
user friendly versions of our two-step estimator, depending on whether one want to use a
first-step consistent estimator of ✓0 or only of ⌫(✓0 ). Note that, this section is generally valid
for estimating equations and their linearizations, irrespective of the fact that these estimating
equations can be seen as first order conditions provided by an extremum estimator, as in our
leading example studied in Section 2.
3.1
The general result
Linear approximations will be considered in some neighborhood @("), " > 0, of the true unknown
value
@(") = ✓ 2 Rp : ✓ ✓0 < " ⇢ ⇥.
Note that, the existence of such " is tantamount to the maintained assumption that the true
unknown value ✓0 belongs to the interior of the parameter space.
In order to extend the results of Robinson (1988), we first characterize our benchmark
estimator ✓ˆT as the solution of some just-identified estimating equations. For sake of generality,
we maintain some high level assumptions about these estimating equations, although they
would in general be implied by more primitive assumptions, as seen in our leading example of
extremum estimation in section 2.
Assumption B1: fT (✓) = qT [✓, ⌫(✓)] is a p-vector valued random variable such that:
(i) fT has a zero ✓ˆT = ✓0 + oP (1),
(ii) For some " > 0, the functions of ✓: ⌫(✓), fT (✓) and
entiable on @("), for any given ✓⇤ in @(").
(iii) FT (✓0 ) = F + oP (1), where FT (✓) =
@fT (✓)
@✓ 0
@qT
@⌫ 0
[✓, ⌫(✓⇤ )] are continuously di↵er-
and F is non-singular.
Under standard regularity conditions (see appendix), the non-singular matrix F can obviously be written as
@q1 [✓0 , ⌫(✓0 )] @q1 0
@⌫
F =
+
[✓ , ⌫(✓0 )] 0 (✓0 ),
0
0
@✓
@⌫
@✓
0
for some population estimating equations q1 [✓, ⌫(✓)] with ✓ the only zero of q1 [✓, ⌫(✓)].
Assumption B2: q1 [✓, ⌫(✓)] = 0 , ✓ = ✓0 .
We are interested in partially linear approximations of the estimating function around some
consistent initial estimator ✓˜T . Thus, let us define
@qT
@⌫
h̃T (✓) = qT [✓, ⌫(✓˜T )] +
[✓, ⌫(✓˜T )] 0 (✓˜T )(✓
0
@⌫
@✓
14
✓˜T ).
Note that h̃T (✓) provides alternative estimating equations that also locally identify ✓ since, with
obvious notations (and under standard regularity conditions), a solution ✓ = ✓T⇤ of gT (✓) = 0
will converge towards a solution ✓ = ✓¯ of the population equations
q1 [✓, ⌫(✓0 )] +
@q1
@⌫
[✓, ⌫(✓0 )] 0 (✓0 )(✓
0
@⌫
@✓
✓0 ) = 0.
With a genuine linearization of fT (✓) (not only a partial one), Robinson’s Theorem 2 shows
that a zero of h̃T (✓) is, in a sense, asymptotically equivalent to ✓ˆT . With a partial linearization,
we cannot maintain such a claim since we may only have local identification and not global
identification. That is, there may exist some ✓¯ 6= ✓0 such that, with obvious notations,
¯ ⌫(✓0 )] + @q1 [✓,
¯ ⌫(✓0 )] @⌫ (✓0 )(✓¯
q1 [✓,
0
@⌫
@✓0
✓0 ) = 0
even though ✓ = ✓0 is the only solution of
q1 [✓, ⌫(✓)] = 0.
To avoid such a perverse situation, we have to slightly penalize our (partially) linearized
sequence by defining:
2
@qT
˜T )] @⌫ (✓˜T )(✓ ✓˜T ) + ↵T ✓ ✓˜T ep ,
[✓,
⌫(
✓
(16)
h̃PT (✓) = qT [✓, ⌫(✓˜T )] +
@⌫ 0
@✓0
for a real sequence ↵T going slowly to infinity, where ep stands for the p-dimensional vector
whose components all equal 1. More precisely, our extension of Robinson’s result can be stated
as follows:
Proposition 3.1: Under standard regularity conditions detailed in the appendix: under
Assumption B1, if ✓˜T is a consistent estimator of ✓0 such that ✓˜T ✓0 = oP (1/↵T ) with
limT =1 ↵T = 1, then for any zero ✓˜P of h̃P (✓) in (16)
T
✓ˆT
T
✓˜TP = OP
✓
↵T
✓ˆT
✓˜T
2
◆
.
Proposition 3.1 is a generalization of Theorem 2 in Robinson (1988). When there is no
function ⌫, our Assumptions B1 and B2 exactly match those of Robinson (1988). However,
there is a price to pay for a linearization that is only partial and thus takes a penalty term
2
↵T going to infinity. Fortunately, the penalty term ↵T ✓ˆT ✓˜T , which shows up in the rate
of convergence, will more often than not have a very minor impact for the use of Proposition
3.1. As a matter of fact, Proposition 3.1. will often be applied to state that, when the initial
estimator ✓˜T is root-T consistent, the two estimators ✓ˆT and ✓˜TP are first order asymptotically
equivalent. This conclusion
is indeed warranted insofar as we pick a penalty rate ↵T going to
p
infinity slower than T . However, the choice of the tuning parameter is more
p constrained if
one wants to use an initially consistent estimator ✓˜T converging slower than T . Exactly as in
15
the case of Robinson (1988), the conclusion of asymptotic equivalence between ✓ˆT and ✓˜TP takes
anyway an initial estimator ✓˜T converging faster than T 1/4 . But, on top of that, if the rate of
convergence of ✓˜T is, say, T (1/4)+" , " > 0, (resp T (1/4) log(T )), the wished asymptotic equivalence
will be warranted only for a slowly diverging penalty rate ↵T like T " (resp. log[log(T )] ). It is
worth noting a tight similarity between the choice of this tuning parameter ↵T and the choice of
the number k(T ) of iterations in iterative procedures like generalized backfitting in Pastorello
et al. (2003) or MBP in Fan et al. (2015). As it can be seen, for instance, on the bottom of
page 465 in Pastorello et al. (2003), k(T ) must go to infinity faster than log(T ) and, in finite
samples, the size of the needed k(T ) is inversely related to the strength of the contraction in
the contraction mapping argument at stake for convergences of the iterations.
3.2
A couple of two-step efficient estimators
Our two-step efficient estimator is a direct extension of Robinson (1988), replacing the complete
linearization by a partial one, and is defined as a zero ✓˜TP of the estimating equations
h̃PT (✓) = qT [✓, ⌫(✓˜T )] +
@qT
@⌫
[✓, ⌫(✓˜T )] 0 (✓˜T )(✓
@⌫ 0
@✓
✓˜T ) + ↵T ✓
✓˜T
2
ep .
These estimating equations provide an efficient estimator by contrast with the naive twostep strategy that would only solve the equations
qT [✓, ⌫(✓˜T )] = 0.
with iteration on these equations a possibility. Up to the penalty term (only used to enforce
consistency), the di↵erence between qT [✓, ⌫(✓˜T )] and h̃PT (✓) is the introduction of the firstorder correction through partial linearization. However, there is an obvious way to make
the estimating equations h̃PT (✓) even more computationally friendly by making the correction
term linear in the unknown parameters ✓; that is, rather, by solving the following estimating
equations:
hT (✓) = qT [✓, ⌫(✓˜T )] +
(1)
@qT ˜
@⌫
[✓T , ⌫(✓˜T )] 0 (✓˜T )(✓
0
@⌫
@✓
✓˜T ) + ↵T ✓
✓˜T
2
ep .
(17)
(1)
The di↵erence between h̃PT (✓) and hT (✓) is that we have also plugged in the first-step
consistent estimator ✓˜T to replace the non-awkward occurrence of the parameters ✓ in the
complete Jacobian matrix. The great thing with this simplifying modification is that it does
not impair the general equivalence result of proposition 3.1
Theorem 3.1: Under standard regularity conditions detailed in appendix: Under assumptions B1 and B2, if ✓˜T is a consistent estimator of ✓0 such that ✓˜T ✓0 = oP (1/↵T ) with
(1)
(1)
limT =1 ↵T = 1, then for any zero ✓T of hT (✓) in (17)
✓
(1)
ˆ
✓T ✓T = OP ↵T ✓ˆT ✓˜T
16
2
◆
.
Theorem 3.1 implies that the previous discussion about the asymptotic efficiency of ✓˜TP ,
(1)
deduced from Proposition 3.1, applies similarly to efficiency of ✓T .
(1)
While ✓T is obviously the most computationally friendly two-step estimator when we have
at our disposal a first-step consistent estimator ✓˜T , it may be a shame to require the use of such
an estimator when, after all, our only trouble is to properly deal with the awkward parameter
(2)
occurrences ⌫(✓). We now propose an alternative two-step efficient estimator ✓T that only
requires knowledge of a first-step consistent estimator ⌫˜T of ⌫(✓0 ), and not the knowledge of a
consistent estimator ✓˜T of the complete parameter vector ✓0 . In applications, it may be typically
the case that only a sub-vector of the parameters of interest ✓ can be consistently estimated in a
first-step. Of course, the price to pay for this additional extension of Robinson (1988) will be to
give up the computational simplification brought by the change from the estimating equations
(1)
h̃PT (✓) to hT (✓) (change from Proposition 3.1 to Theorem 3.1). By definition, if we don’t have
such thing as a first-step estimator ✓˜T , we cannot plug it in to simplify the equations.
However, we will be able to derive an alternative two-step efficient estimator through the
estimating equations defined by
(2)
hT (✓) = qT [✓, ⌫˜T ] +
@qT
[✓, ⌫˜T ].(⌫(✓)
@⌫ 0
⌫˜T ) + ↵T k⌫(✓)
⌫˜T k2 ep .
(18)
Theorem 3.2: Under standard regularity conditions detailed in the appendix: under assumptions B1 and B2, if ⌫˜T is a consistent estimator of ⌫(✓0 ) such that k˜
⌫T ⌫(✓0 )k = oP (1/↵T )
(2)
(2)
with limT =1 ↵T = 1, then for any zero ✓T of hT (✓) in (18)
✓
◆
2
(2)
ˆ
ˆ
✓T ✓T = OP ↵T ⌫(✓T ) ⌫˜T
.
Theorem 3.2 implies that the previous discussion about the asymptotic efficiency of ✓˜TP and
(1)
✓T , deduced from Proposition 3.1 and Theorem 3.1, respectively, applies similarly to efficiency
(2)
of ✓T . The main di↵erence is that the leading rate of convergence is now the one of the
estimator ⌫˜T . It is also worth noting that the idea of the proof of Theorem 3.2 can be applied
even when plugging in a first-step consistent estimator ✓˜T to replace part of or all components
of the first occurrence of ✓ in the Jacobian term. In particular, the two simplifying ideas of
Theorems 3.1 and 3.2 can be used simultaneously.
3.3
Practical implications:
In this subsection, we set the focus on the simplest case where both the benchmark efficient
estimator ✓ˆT and the initial estimators ✓˜T or ⌫˜T are all root-T consistent.
When ✓ˆT is defined as the solution of
qT [✓ˆT , ⌫(✓ˆT )] = 0,
we propose two more user-friendly estimators, both associated with a sequence ↵T of tuning
parameters.
17
(1)
First, ✓T defined as solution of:
@qT ˜
@⌫
(1)
(1)
qT [✓T , ⌫(✓˜T )] +
[✓T , ⌫(✓˜T )] 0 (✓˜T )(✓T
0
@⌫
@✓
(1)
✓˜T ) + ↵T ✓T
2
✓˜T
ep = 0.
(19)
(2)
Second, ✓T defined as solution of:
(2)
qT [✓T , ⌫˜T ] +
@qT (2)
(2)
[✓ , ⌫˜T ].(⌫(✓T )
@⌫ 0 T
(2)
⌫˜T ) + ↵T ⌫(✓T )
⌫˜T
2
ep = 0.
(20)
Recall that several variants are possible, depending upon what part of the first-step estimator
is used for the penalty term and/or for computing the derivative @qT /@⌫ 0 .
For practical choice of the tuning parameter sequence ↵T , the two golden rules are
p as follows.
First, for sake of asymptotic efficiency, ↵T must go to infinity strictly slower than T ; Second,
the fact that ↵T goes to infinity is only useful to ensure consistency (see Step 1 in the proof of
Proposition 3.1).
In many circumstances, consistency will be warranted even without the penalty, that is, with
↵T = 0. This, in particular, paves the way for many efficient two-step extremum estimators
as exemplified in sections 2.2 and 2.3. Generally speaking, when consistency is not an issue,
Theorem 2.1 states asymptotic efficiency of two-step extremum estimators ✓ˆText computed as
solutions of
⇢
@QT [✓, ⌫˜T ]
1
ext
ˆ
✓T = arg max QT [✓, ⌫˜T ] +
. [⌫(✓) ⌫˜T ]
[⌫(✓) ⌫˜T ]0 JT (✓) [⌫(✓) ⌫˜T ] . (21)
✓
@⌫ 0
2
Moreover, the proof of Theorem 2.1 shows that the dependence on ✓ of the weighting matrix
JT (✓) can be overlooked in computing the first order conditions and then, up to the penalty
term, the first order conditions for (21) are very similar to (19). Sections 2.2 and 2.3 display
user friendly closed form formulas for the weighting matrix JT (✓) that do not involve any second
derivatives. However, it is important to keep in mind that consistency is not always warranted,
and then, the only solution is the introduction of the penalty term in first order conditions
leading to (19) or (20).
4
4.1
Additive decomposition of Extremum Criterion
Efficient two-step Estimation via Margin Targeting
There exist many interesting situations in economics and finance where the extremum criterion
takes the additively separable form
QT [✓, ⌫(✓)] = Q1T [✓1 ] + Q2T [✓2 , ⌫(✓)],
(22)
where ✓ = (✓10 , ✓20 )0 , ⌫(✓) = ✓1 2 Rp1 , ✓2 2 Rp2 and p1 + p2 = p. This particular structure
for QT [✓, ⌫(✓)] includes many nonlinear time series models, such as, the Dynamic Conditional
Correlations (DCC-GARCH) model of Engle (2002), the rotated ARCH model of Noureldin
et al. (2014), and many copula models. In these multivariate models ✓1 generally represents the
18
parameters that govern the marginal distributions and ✓2 represent the parameters that govern
the dependence between the di↵erent components. In this framework, ⌫(✓) = ✓1 represents
the additional occurrences of ✓1 that show up in the dependence structure and complicate
estimation of ✓.
In this setting, a common way of estimating ✓ = (✓10 , ✓20 )0 is the so-called inference from
the margins, where a root-T consistent estimator ✓˜T is obtained by first maximizing Q1T [✓1 ] to
obtain ✓˜1T , which is equivalent to solving the estimating equations
@Q1T [✓˜1T ]
= 0,
@✓1
(23)
✓˜1T then replaces the unknown ✓1 in Q2T [✓2 , ✓1 ] and Q2T [✓2 , ✓˜1T ] is maximized to obtain ✓˜2T ,
which is equivalent to solving
@Q2T [✓˜2T , ✓˜1T ]
= 0.
@✓2
(24)
If (23) and (24) are unbiased estimating equations for ✓0 , in the sense that,
@Q1T [✓1 ]
= 0 () ✓1 = ✓10 ,
T =1
@✓1
@Q2T [✓2 , ✓10 ]
lim
= 0 () ✓2 = ✓20 ,
T =1
@✓2
lim
0
0
✓˜T = (✓˜1T
, ✓˜2T
)0 is generally a root-T consistent estimator of ✓0 .
While computationally simple, the estimator ✓˜T is inefficient, which is seen by noting that
the efficient estimator ✓ˆT , the maximizer of QT [✓, ⌫(✓)], solves the estimating equations

q1T [✓, ⌫(✓)]
qT [✓, ⌫(✓)] =
,
q2T [✓, ⌫(✓)]
where
@Q1T [✓1 ] @Q2T [✓2 , ⌫(✓)]
+
,
@✓1
@⌫
@Q2T [✓2 , ⌫(✓)]
q2T [✓, ⌫(✓)] =
.
@✓2
q1T [✓, ⌫(✓)] =
Computationally simple and efficient estimators can be obtained in this setting using the two(1)
(2)
step estimators ✓T and ✓T , defined in Section 3.2 as the solutions to the estimating equations
(1)
(2)
(1)
(2)
0 = hT (✓) and 0 = hT (✓) (with hT (✓) and h̃T (✓) given in (19) and (20) respectively), and
first-step estimator ✓˜T defined by estimating equations (23) and (24).
(1)
(2)
Obtaining ✓T and ✓T when QT [✓, ⌫(✓)] is additively separable following (22) then requires
19
(1)
(2)
specializing the definitions of hT (✓) and hT (✓). To this end, for
"
#
"
#
(1)
(2)
h1T (✓)
h1T (✓)
(1)
(2)
hT (✓) =
, hT (✓) =
,
(1)
(2)
h2T (✓)
h2T (✓)
(1)
(1)
we have that ✓T , defined as the solution to 0 = hT (✓), solves
(1)
(1)
@Q1T [✓1T ] @Q2T [✓2T , ✓˜1T ] @ 2 Q2T [✓˜2T , ✓˜1T ] (1) ˜
(1)
+
+
(✓1T ✓1T ) + ↵T ✓T
@✓1
@⌫
@⌫@⌫ 0
(1)
2
@Q2T [✓2T , ✓˜1T ] @ 2 Q2T [✓˜2T , ✓˜1T ] (1) ˜
(1) (1)
(1)
˜T ep
0 = h2T (✓T ) =
+
(✓
✓
)
+
↵
✓
✓
1T
T
2
1T
T
@✓2
@✓2 @⌫ 0
(1)
(1)
0 = h1T (✓T ) =
(2)
✓˜T
2
ep1(25)
(26)
(2)
and ✓T , defined as the solution to 0 = hT (✓), solves
(2)
(2)
(2)
2
@Q1T [✓1T ] @Q2T [ ✓2T , ✓˜1T ] @ 2 Q2T [✓2T , ✓˜1T ] (2) ˜
(2)
+
( ✓1T ✓1T ) + ↵T ✓1T ✓˜1T ep1(27)
0
@✓1
@⌫
@⌫@⌫
(2) ˜
(2) ˜
2
2
@Q2T [✓2T , ✓1T ] @ Q2T [✓2T , ✓1T ] (2) ˜
(2) (2)
(2)
0 = h2T (✓T ) =
+
(✓1T ✓1T ) + ↵T ✓1T ✓˜1T ep2 ,
(28)
0
@✓2
@✓2 @⌫
p
for some sequence ↵T going to infinity slower than T .
(1)
Obviously, solving (25) and (26) (respectively, (27) and (28)) to obtain ✓T (respectively,
(2)
(1)
(2)
✓T ) is more computationally involved than the estimator ✓˜T . However, both ✓T and ✓T share
with ✓˜T the convenient feature that the cumbersome occurrence of ✓1 in Q2T [✓2 , ✓1 ] never shows
up as an unknown parameter in the estimating equations, which makes our two-step efficient
estimator computationally friendly in comparison with the brute force efficient estimator ✓ˆT .
This simplification of the estimating equations is also shared by the MBP estimator proposed
in SFK. When QT [✓, ⌫(✓)] is additively separable following (22), the MBP algorithm takes as
(k)
its starting value ✓˜T and defines a sequence of iterative estimators ✓ˆT , k > 1, by solving
(2)
(2)
0 = h1T (✓T )
(k+1)
(k) (k)
@Q1T [✓ˆ1T ] @Q2T [✓ˆ2T , ✓ˆ1T ]
+
,
@✓1
@✓1
(k+1) (k)
@Q2T [✓ˆ2T , ✓ˆ1T ]
0 =
.
@✓2
0 =
While each iteration of the MBP procedure is computationally simpler than the second-step
of the penalized two-step estimators, the price to pay for this simplicity is two-fold: one, to
achieve efficiency we require k ! 1, possibly according to a tuning parameter k = k(T ), and
two, convergence of the MBP iterations requires the existence of a local contraction mapping
condition, often called an information dominance condition.
If the information dominance condition is nearly unsatisfied, the MBP iterations converge
(k)
very slowly, and if this condition is not satisfied ✓ˆT does not converge. To deal with such
situations FPR propose a modification of the MBP estimator in SFK that regains a portion
of the information associated with the occurrence of ✓2 in Q2T [✓2 , ✓1 ] neglected by the original
20
(k)
MBP scheme. Consequently, FPR define this alternative MBP estimator ✓˜T as the solution
to the following estimating equations,
(k+1)
(k+1) (k)
@Q1T [✓˜1T ] @Q2T [✓˜2T , ✓˜1T ]
0 =
+
,
@✓1
@✓1
(k+1) (k)
@Q2T [✓˜2T , ✓˜1T ]
0 =
.
@✓2
(29)
(30)
Note that this estimator is nothing but the MBP estimator conformable to the general definition
(4).
It is straightforward to compare the computational burden associated with the MBP esti(1)
mator in (29), (30) and the two-step penalized estimator ✓T (dubbed P-TS1 ), as well as the
(1)
additional two-step estimator ✓T (P ) (dubbed TS1 ) that arises from neglecting the penalty terms;
(1)
i.e., the TS1 estimator ✓T (P ) solves the estimating equations (25, 26), but with ↵T = 0.1 Firstly,
comparing the MBP estimator and TS1 (respectively, P-TS1 ), the only di↵erence between the
two estimators is that TS1 (respectively, P-TS1 ) entails some minor computational burden as(1)
sociated with the introducing of a linear function of ✓1T (this statement holds up to the penalty
term for P-TS1 ). This tiny additional complexity is the price to pay to get efficiency in two
steps instead of fishing for the limit of an iterative procedure, which, as stated above, may
require many iterations depending on the strength of the local-contraction mapping.
However, when the local contraction mapping is strong, the MBP procedure of SFK is
the simplest from a computational standpoint. As the required contraction mapping condition
becomes weaker, the MBP estimator becomes more computationally burdensome.2 In contrast,
the two-step procedures discussed herein do not require a contraction mapping condition and
can therefore yield consistent and efficient estimators in situations where this condition is
violated.
(2)
In comparison with the aforementioned estimators, the penalized two-step estimator ✓T
(2)
(dubbed P-TS2 ), and the corresponding version ✓T (P ) (dubbed TS2 ) that neglects the penalty
function, incurs additional computational complexity because ✓2 occurs within the partial Hessian term in the estimating equations. However, in this setting, the P-TS2 (and TS2 ) estimator
is unique in that it only requires a consistent first-step estimator for ✓10 , and not for ✓20 . In the
framework of estimation from the margins, this advantageous property of TS2 (and P-TS2 )
can be interpreted as follows. In many multivariate models, ✓1 can simply be estimated from
the margins and is numerical stable. In contrast, estimation of the dependence parameters
✓2 is often tricky and numerically unstable. Indeed, this is a primary reason why (unconditional) variance targeting, as initially proposed by Engle and Mezrich (1996), became popular
in the estimation of multivariate GARCH models, with similar reasoning leading researchers
to contemplated correlation targeting in estimation of GARCH-DCC models. From a targeting standpoint, the P-TS2 (TS2 ) estimator first obtains a simple estimate ✓˜1T of ✓10 from the
margins, then uses ✓˜1T via a ”margin targeting” procedure whereby the second-step of the
1
Note that, from Proposition 3.1 and Theorems 3.1 and 3.2, when consistent the two-step estimators that
disregards the penalty term will also be asymptotically efficient.
2
This statement also holds for the MBP estimator proposed in FPR.
21
estimation procedure is stabilized by targeting the consistent marginal parameter estimates.
In contrast to (unconditional) variance targeting, P-TS2 (and TS2 ) does not incur an efficiency loss associated with margin targeting. More importantly, P-TS2 (and TS2 ) need not
maintain the problematic assumption in unconditional variance targeting on the existence of
higher order unconditional moments, which is required in order to for variance targeting to
yield an asymptotically normal estimator of the unconditional variance.
4.2
Bivariate Gaussian Copula Models
In the following subsection, we illustrate the above discussion between the di↵erent estimation
procedures using a Gaussian Copula model. The Bivariate Gaussian copula model has been
extensively studied in statistics and economics, see, e.g., Joe (1997), Song (2000), among others,
and is often used in empirical analysis.
Assume our goal is to estimate the parameters governing the distribution of yi = (yi,1 , yi,2 )0 .
Denoting the marginal distribution of yi,j as Fj (·; ↵j ), where ↵j is a vector of unknown parameters, the joint distribution can be constructed using a copula function C(u1 , u2 ; ⇢), where ⇢
denotes the copula dependence parameter. In what follows, we assume yi = (yi,1 , yi,2 )0 follows
a bivariate Gaussian copula with cumulative distribution function (CDF)
C(F1 (yi,1 ; ↵1 ), F2 (yi,2 ; ↵2 ); ⇢) =
⇢(
1
(F1 (yi,1 ; ↵1 )),
1
(F2 (yi,2 ; ↵2 ))),
(31)
where ⇢ (·) is the bivariate Gaussian cumulative distribution function with correlation parameter ⇢ and (·) is the standard normal CDF. Denote by c(F1 (yi,1 ; ↵1 ), F2 (yi,2 ; ↵2 ); ⇢) the copula
density derived from equation (31). For (u1 , u2 )0 2 (0, 1)2 , Song (2000) demonstrates that the
density of the bivariate Gaussian copula is
✓
◆
1
⇢(z12 + z22 ) 2⇢(z1 · z2 )
c(u1 , u2 ; ⇢) = p
exp
,
2(1 ⇢2 )
1 ⇢2
1
where zj =
(uj ) for j = 1, 2.
Let fj (yi,j ; ↵j ) denote the marginal density of yi,j and define ✓1 = (↵10 , ↵20 )0 , ✓2 = ⇢, with
✓ = (✓10 , ✓2 )0 . Inference for ✓ in the Bivariate Gaussian copula model can be carried out using
maximum likelihood, with corresponding log-likelihood function
QT [✓, ⌫(✓)] =
T X
2
X
i=1 j=1
log(fj (yi,j ; ↵j ))
T
log(1
2
⇢2 )
⇢
2(1
⇢2 )
(⇢A(✓1 )
2B(✓1 )).
(32)
PT
PT
2
2
Herein, A(✓1 ) =
i=1 [zi,1 (↵1 ) + zi,2 (↵2 ) ], B(✓1 ) =
i=1 zi,1 (↵1 )zi,2 (↵2 ), and zi,j (↵j ) =
1
(Fj (yi,j ; ↵j )) for j = 1, 2. The likelihood in (32) is separable and we denote the two pieces
Q1T [✓1 ] =
T X
2
X
log(fj (yi,j ; ↵j )), and Q2T [✓2 , ⌫(✓)] =
i=1 j=1
where, again, ⌫(✓) = ✓1 .
22
T
⇢
log(1 ⇢2 )
(⇢A(✓1 ) 2B(✓1 )),
2
2(1 ⇢2 )
4.2.1
Estimators of ✓
Depending on the specification of the marginals fj (·; ↵j ), maximizing QT [✓, ⌫(✓)] to obtain the
Maximum Likelihood estimator (MLE) ✓ˆT can be difficult. In these cases a simple two-step
estimation approach, the so-called inference from margins (IFM) approach, is often used to
estimate ✓ (see, e.g., Shih and Louis (1995), Joe (1997) and
P Patton
P (2009) for examples and
discussion). The IFM approach first maximizes Q1T [✓1 ] = Ti=1 2j=1 log(fj (yi,j ; ↵j )) to obtain
0
0
✓˜1T = (˜
↵1T
,↵
˜ 2T
)0 , defined as the solution to
!
Pn
@f1 (yi,1 ;↵1 )
1
@Q1T [✓1 ]
i=1 f1 (yi,1 ;↵1 )
@↵1
0=
= Pn
.
@f2 (yi,2 ;↵2 )
1
@✓1
i=1 f (y ;↵ )
@↵
2
i,2
2
2
Next, the unknown ✓1 in Q2T [✓2 , ✓1 ] is replaced with ✓˜1T and Q2T [✓2 , ✓˜1T ] = T2 log(1 ⇢2 )
⇢
(⇢A(✓˜1T ) 2B(✓˜1T )) is maximized to obtain ✓˜2T = ⇢˜T , defined as the solution to
2(1 ⇢2 )
0=
@Q2T [✓˜1T , ✓˜2T ]
T⇢
=
@✓2
1 ⇢2
1
(⇢A(✓˜1T )
(1 ⇢2 )2
(1 + ⇢2 )B(✓˜1T )).
It is clear from this decomposition that the IMF estimator disregard the information about ✓1
contained in
!
n
@B(✓1 )
1)
X
⇢ @A(✓
2
@Q2T [✓2 , ✓1 ]
⇢
@↵1
@↵1
=
.
@A(✓1 )
@B(✓1 )
2
@✓1
1
⇢
⇢
2
@↵2
@↵2
i=1
From the above definitions, we see that the efficient MBP and penalized two-step estimators obtain efficiency by adding back, in di↵ering combinations, terms associated with
@Q2T [✓2 , ✓1 ]/@✓1 . MBP accomplishes this task by adding back @Q2T [✓2 , ✓1 ]/@✓1 to the estimating equations for ✓1 and iterating over the cumbersome occurrences of ✓1 (and ✓2 , depending on
(1)
the precise MBP method). On the other hand, the penalized two-step estimator ✓T (previously
dubbed P-TS1 ) linearizes @Q2T [✓2 , ✓1 ]/@✓1 , with respect to the cumbersome occurrence of ✓1 ,
around the consistent estimator ✓˜1T , and targets the second-step estimators using the initially
(2)
consistent ✓˜T . The penalized two-step estimator ✓T (previously dubbed P-TS2 ) is similar to
P-TS1 but only penalizes the estimating equations with respect to the margins estimator ✓˜1T .
Both two-step approaches have the same asymptotic distribution, but can behave di↵erently in
finite samples.
In comparison with the two-step procedures, the critical regularity condition needed for
the MBP estimator to be efficient is the satisfaction of a local contraction mapping condition,
also termed the information dominance condition. However, in the bivariate Gaussian copula
model, simulation evidence in SFK and Liu and Luger (2009) demonstrate that the MBP
approach can behave poorly if there is even moderate correlation. Intuitively, this phenomena
is present because as ⇢ increases the portions of the estimating equations that MBP iterates
over become more informative for estimating the parameters. For ⇢ large enough the MBP
algorithm neglects too much information and yields an inconsistent estimator.
23
4.2.2
Example: Exponential Marginals
In this subsection we compare the finite sample properties of the MBP approach of SFK and
four di↵erent efficient two-step procedures: the penalized two-step estimator P-TS1 , the nonpenalized counterpart to P-TS1 given by TS1 , the partially penalized two-step estimator P-TS2 ,
and the non-penalized counterpart to P-TS2 given by TS2 . Data for the exercise is generated from the Gaussian copula in the situation where the marginal densities are exponential:
fj (yi,j ; ↵j ) = ↵j exp( ↵j yi,j ), ↵j > 0, j = 1, 2.
In particular, the simulation study compares the e↵ects of the correlation parameter and
sample size on the various estimators. For the simulation study we set ↵1 = .1, ↵2 = 1
and consider three di↵erent values for the correlation parameter ⇢ = .75, .95, .985. Across the
three values of ⇢ we consider three di↵erent sample sizes T = 100, 200, 300. For each T and ⇢
combination we create 1,000 synthetic samples.
It is important to note that for ⇢ greater than approximately .95 the information dominance
condition associated with the proposed MBP procedure is no longer satisfied. Therefore, at
high levels of correlation we expect the finite sample properties of the MBP estimator to be
poor in comparison with the various two-stage estimators.
The estimators are compared in terms of their means, mean squared error (MSE) and mean
absolute error (MAE), across the di↵erent sample sizes. We define convergence for the MBP
algorithm as the maximum absolute di↵erence across the parameters being less than 1.0e 05
for two or more successive iterations. Tables 1 to 3 report the averages over the 1,000 synthetic
samples for the mean, MSE and MAE across the three correlation values ⇢ = {.75, .95, .985}.
For the penalized two-step estimators the penalty term is taken proportional to T 1/4 .
For low values of the correlation parameter the MBP algorithm and the efficient two-step
estimators are very similar. However, as the correlation parameter increases, the penalized
two-step methods give smaller MSEs and MAEs than the MBP estimator and non-penalized
two-step estimator. With high correlation values and larger sample sizes the MBP algorithm
encounters difficulty in estimation since the matrix driving the updates does not fulfill the IDC.
It is important to point out that the same behavior is not found in the two-stage and penalized
two-stage estimates, which perform well even for ⇢ = .985.
The various combinations of penalized and non-penalized two-step estimators all deliver
stable parameter estimates with good finite sample properties. However, the fully penalized
estimator P-TS1 does seem to have a slight edge over the other estimators in terms of performance. The small impact of the penalty term in this situation is very easy to interpret:
for copula models the IFM procedure often provides accurate starting values, and therefore
the need to penalize is drastically reduced.3 In other words, in the copula case, the two-step
procedure merely ensures efficiency and penalization seems not to be required.
5
Efficient two-step estimation with Implied States
In this section we analyze situations where ✓0 is determined by the law of motion governing a
latent stochastic process of interest {Yt⇤ : t 1}. The latent state variables Yt⇤ are unobservable
3
Recall, the penalty term is needed to rule out any perverse solutions to the estimating equations, which can
exist because of the partial linearization.
24
to the econometrician, but are related to observed data Yt through a function h[·, ⌫ 0 ], known
up to the unknown parameters ⌫ 0 = ⌫(✓0 ), according to the relationship
Yt = h[Yt⇤ , ⌫ 0 ].
We are only interested in situations where Y ⇤ 7! g[Y ⇤ , ⌫] is one-to-one for any ⌫, which implies
that, if ⌫ 0 was known, Yt⇤ could be directly obtained by inverting h[·, ⌫ 0 ]; i.e.,
Yt = h[Yt⇤ , ⌫ 0 ] () Yt⇤ = g[Yt , ⌫ 0 ].
(33)
When ⌫(✓0 ) is unknown, equation (33) defines the implied state (variable) Yt⇤ (✓) = g[Yt , ⌫(✓)].
As has been noted by several authors, such as, e.g., Renault and Touzi (1996), and Pastorello
et al. (2003), the setup in (33) covers many interesting applications in economics and finance.
However, estimation of ✓0 is often complicated by the nature of the function h[·, ⌫] and the
difficulties encountered when transforming the estimation problem from one based on latent
states Yt⇤ , to one based on implied states g[Yt , ⌫(✓)].
In what follows, we demonstrate that the efficient penalized two-step estimator can often
be used to obtain consistent and efficient estimators for ✓0 in models with implied states. In
particular, we focus on the use of implied states in GMM, so-called, Implied States GMM, and
in likelihood models with latent states. A comparison with existing estimation approaches in
these settings is also given.
5.1
Implied States GMM
Pan (2002) uses the terminology Implied States GMM (IS-GMM) to describe GMM estimation
in the context of option pricing models with latent variables. More specifically, the IS-GMM
estimator of Pan (2002) uses observed option price data to back-out, through an option pricing
formula, the latent state variables driving the price process.
Formally, we are interested in analyzing a model with true parameter ✓0 , defined as the
unique zero of a vector of moment conditions derived from the law of motion for Yt⇤ :
E[
⇤
(Yt⇤ , ✓)] = 0 () ✓ = ✓0 .
(34)
Clearly, GMM estimation from (34) is not feasible since Yt⇤ is unobservable. Implementation of
GMM in this setting can, however, be carried out by substituting the implied states, say, g[Yt , ✓],
which are obtained by inverting Yt = h[Yt⇤ , ✓] to get Yt⇤ = g[Yt , ✓], into (34). The existence of
the one-to-one relationship in (33) is common in many arbitrage-based asset pricing models.
For instance, in options pricing Yt may be the observed option price, Yt⇤ can represent the latent
variables driving the price process and h[Yt⇤ , ✓] will be the pricing formula linking Yt and Yt⇤ .
Plugging the implied states g[Yt , ✓] into (34) yields
E[ (Yt , ✓, ⌫(✓))] = 0, with
(Yt , ✓, ⌫(✓)) =
⇤
(g[Yt , ⌫(✓)], ✓).
(35)
The first occurrence of ✓ within (35) represents the original occurrences of ✓ in the moment
conditions represented by the latent data, while ⌫(✓) represents the occurrences of ✓ in the
25
implied states. This later occurrence of ⌫(✓) in (·) is generally computationally cumbersome
in comparison with the former occurrence of ✓ in (·).
When the moment conditions in (35) are overidentified, we take as our extremum criterion
QT [✓, ⌫(✓)] the efficient two-step GMM criterion:
QT [✓, ⌫(✓)] =
¯ T [✓, ⌫(✓)]0 W
1
T
(✓˜T ) ¯ T [✓, ⌫(✓)],
P
where ¯ T [✓, ⌫(✓)] = T1 Tt=1 (Yt , ✓, ⌫(✓)), ✓˜T is a preliminary consistent estimator, and WT 1 (✓˜T )
p
is a consistent estimator for the long-run variance matrix of T ¯ T [✓0 , ⌫(✓0 )]. The efficient estimator ✓ˆT can then be defined as the (unique) zero of the estimating equations
 ¯
@ T [✓, ⌫(✓)] @⌫ 0 (✓) @ ¯ T [✓, ⌫(✓)]
0=
+
@✓
@✓
@⌫
0
WT 1 (✓˜T ) ¯ T [✓, ⌫(✓)].
(36)
Note that, the estimator ✓ˆT uses a consistent estimator for the selection matrix

0
@ [✓0 , ⌫(✓0 )] @⌫ 0 (✓0 ) @ [✓0 , ⌫(✓0 )]
0
(✓ ) = E
+
W 1 (✓0 ).
@✓
@✓
@⌫
In contrast, the simpler IS-GMM estimator ✓TIS defined as the solution of
max
✓
solves the estimating equations
"
¯ T [✓, ⌫(✓˜T )]0 W
@ ¯ T [✓, ⌫(✓˜T )]
0=
@✓
#0
1
T
(✓˜T ) ¯ T [✓, ⌫(✓˜T )]
WT 1 (✓˜T ) ¯ T [✓, ⌫(✓˜T )],
and therefore employs a consistent estimator for the selection matrix

0
0
0
e(✓0 ) = E @ [✓ , ⌫(✓ )] W 1 (✓0 ).
@✓
(37)
When dim( ) > dim(⇥), the selection matrix e(✓0 ) selects p linear combinations of the estimating equations in a suboptimal manner, and so ✓TIS will be inefficient in general. Intuitively,
the inefficiency of ✓TIS is a direct consequence of the estimators disregard for the impact of the
awkward occurrences ⌫(✓) of ✓ on the selection matrix, through the Jacobian matrix.
Unlike the unconditional moment setting described herein, Pan (2002) considers the application of IS-GMM in the context of conditional moment restrictions. However, when it comes
to optimal instruments, the same inefficiency issue will be faced if we overlook components of
the Jacobian matrix associated with the occurrences of ✓ in the implied states.
Besides the above IS-GMM estimators, Pastorello et al. (2003) propose an iterative latent
backfitting estimator that defines estimates ✓˜Tk through the iterations
✓˜Tk+1 = arg max QT [✓, ⌫(✓˜Tk )].
✓
26
Upon convergence, ✓˜Tk solves the estimating equations
 ¯
@ T [✓, ⌫(✓)]
0=
@✓
0
WT 1 (✓˜T ) ¯ T [✓, ⌫(✓)],
(38)
and therefore, similar to ✓TIS , the latent backfitting estimator of Pastorello et al. (2003) is
inefficient when dim( ) > dim(⇥).
An alternative to directly solving (36) and the inefficient estimators that solve (37), (38), is
the penalized two-step estimator developed herein. Clearly, we have at our disposal an initial
consistent estimator ✓˜T of ✓0 . Moreover, ✓˜T can also be used to consistently estimate the optimal
instruments via
"
#0
¯ T [✓˜T , ⌫(✓˜T )] @⌫ 0 (✓˜T ) @ ¯ T [✓˜T , ⌫(✓˜T )]
@
˜
+
WT 1 (✓˜T ).
T ( ✓T ) =
@✓
@✓
@⌫
The existence of a consistent estimator for (✓0 ) allows us to define a new two-step estimator
(1)
that utilizes T (✓˜T ) to simplify the existing two-step estimator ✓T defined in equation (19).
To this end, defining qT [✓, ⌫(✓˜T )] = (✓˜T ) ¯ T [✓, ⌫(✓˜T )], we can obtain a simplified efficient
(1)
two-step estimator ✓T⇤IS , in the spirit of ✓T , by solving
˜
0 = h⇤IS
T (✓) = qT [✓, ⌫(✓T )] +
˜
˜
¯ ˜
˜ @ T [✓T , ⌫(✓T )] @⌫(✓T ) (✓
@⌫
@✓0
T ( ✓T )
✓˜T ) + ↵T k✓
✓˜T k2 .
(1)
(1)
The main di↵erence between ✓T⇤IS and ✓T , is that ✓T requires di↵erentiating ¯ T [✓, ⌫(✓)] with
respect to ⌫(✓) and also the occurrences of ⌫(✓) in T (✓). In other words, for the estimator
✓T⇤IS , T (✓) is calculated once and is not altered thereafter. Efficiency of ✓T⇤IS can be shown by
a direct application of Theorem 3.1.
The two-step estimator ✓T⇤IS is similar to the IS-GMM estimator developed in FPR, and
(k)
defined by the sequence of estimators ✓ˆT , the solutions of
0=
ˆ(k 1) ) ¯ T [✓, ⌫(✓ˆ(k 1) )].
T ( ✓T
T
In comparison, neither the two-step or MBP estimator actively search over the cumbersome
occurrences of ✓ in ⌫(✓). In this way, both approaches share some of the computational simplicity associated with the inefficient estimators ✓TIS and ✓˜Tk , however, in contrast both estimators
retain efficiency (under certain conditions). The main di↵erence between the two approaches is
that the two-step approach directly corrects the information loss associated with not optimizing over the occurrences of ✓ due to ⌫(✓) by forming a consistent estimator of these quantities,
the FPR approach on the other hand only o↵ers this correction as k ! 1, and only if the
required contraction condition is satisfied.4 The price to pay for this two-step procedure is
the additional computational cost induced by the linear term in h⇤IS
T (✓) and the addition of a
penalty to guarantee consistency.
4
See FPR for a precise statement of this contraction mapping condition.
27
5.2
Implied States in Latent Likelihood
Let us now consider the case where the unobservable stochastic process {Yt⇤ : t
from a transition density that is known up to the unknown ✓0 , and let
1} is drawn
P = {f (·|·; ✓) : ✓ 2 ⇥}
denote the family of transition densities indexed by ✓. Denoting the log-likelihood based on
the unobservable latent state variables Yt⇤ by
Q⇤T [✓]
T
1X
=
`(Yt⇤ |Yt⇤ 1 ; ✓), where `(Yt⇤ |Yt⇤ 1 ; ✓) = log(f (Yt⇤ |Yt⇤ 1 ; ✓)),
T t=1
the implied states framework utilizes the relationship Yt = h[Yt⇤ , ⌫ 0 ] to transform the estimation
problem from one based on Yt⇤ and Q⇤T [✓] to one based on Yt . Using the implied states g[Yt , ⌫(✓)],
obtained by inverting (33) at the value ✓, and the Jacobian formula, the infeasible log-likelihood
Q⇤T [✓] is transformed into the feasible log-likelihood
T
T
1X
1X
QT [✓, ⌫(✓)] =
`(g[Yt , ⌫(✓)]|g[Yt 1 , ⌫(✓)]; ✓) +
log |Hy g[Yt , ⌫(✓)]| .
T t=1
T t=1
|Hy g[Yt , ⌫(✓)]| is the absolute value of the Jacobian for Y associated with the map Y 7!
g[Y, ⌫(✓)].
Estimation of ✓0 from QT [✓, ⌫(✓)] is often encountered in estimation of option pricing models,
see, e.g., Renault and Touzi (1996), as well as credit risk models, see, e.g., Duan (1994).
Maximization of QT [✓, ⌫(✓)] is generally much more difficult than would be maximization of
Q⇤T [✓], if such maximization were indeed feasible.
It is clear that directly solving
0 = qT [✓, ⌫(✓)] =
@QT [✓, ⌫(✓)] @⌫ 0 (✓) @QT [✓, ⌫(✓)]
+
@✓
@✓
@⌫
can be cumbersome, as ✓ shows up in several places within QT [✓, ⌫(✓)] and in highly nonlinear
ways. While the two-step procedures discussed herein can be applied to such settings, it is
perhaps more informative to consider precise implementation of these estimators in a relatively
simple example.
5.2.1
Example: Merton Credit Risk Model
To demonstrate the penalized two-step methodology in the situation of implied states likelihood
estimation, we now consider estimation of the parameters in the structural credit risk model of
Merton (1974).
Suppose that the firm’s debt consists of a zero coupon bond with face value B and maturity
date . Letting Vt denote the firm’s unobservable market value at time-t, the firm’s observable
equity price can be interpreted as an European call option written on the firm’s market value
28
with strike price B and maturity ; i.e.,
S ⌘ max[V
B, 0].
(39)
From (39) the observed equity prices S0 , ..., ST can be interpreted as option prices written on
the firm’s unobservable market values V0 , ..., VT .
In the simplest case, the firm’s unobservable market value is described as a Geometric
Brownian Motion:
dVt
= µdt + dWt ,
(40)
Vt
where Wt is a standard Brownian motion. Equation (40) allows us to write the conditional
likelihood of the sample path (V1 , V2 , ..., VT ) given some initial value V0 and historical parameters
(µ, ). The conditional log-likelihood function of the unobserved asset values is then given by
Q⇤T [µ,
2
]=
1
ln(2⇡
2
2
)
T
1 X ln(Vt /Vt 1 )
2T t=1
(µ
2
1
2
2
)
2
n
1X
ln Vt ,
T t=1
see, e.g., Duan (1994, 2000) and FPR for a discussion. Unfortunately, maximum likelihood
estimation of (µ, ) from Q⇤T [µ, 2 ] is not feasible since the sample path (V1 , V2 , ..., VT ) is unobserved.
However, when the dynamics of the firm’s market value are described by (40), the observable
equity values can be related to the unobservable firm values through the Black and Scholes
option pricing formula:
p
St = Vt (dt ) B exp( r(
t)) (dt
t),
(41)
p
where dt ( 2 ) = ln(Vt /B) + (r + 12 2 )(
t)/
t, (·) is the standard normal CDF and r
is the risk-free interest rate assumed to be deterministic and time-invariant. Letting g[·, 2 ]
denote the inverse of the Black and Scholes option pricing formula, the unobserved firm values
are related to the observed equity prices through
2
Vt = g[St ,
],
which can be obtained, at least numerically, from equation (41) and a given value of 2 . Technically g[·, 2 ] depends on t through the time-to-maturity (
t), however, we eschew this
dependence in favor of notational simplicity.
Therefore, even though Vt is unobserved, if 2 were known its value could be imputed from
Vt = g[St , 2 ] for each t = 1, ..., T . Given this fact, using Vt = g[St , 2 ] and the Jacobian
formula, we transform the log-likelihood from one based on Vt to one based on St . Following
arguments in Duan (1994), the conditional log-likelihood based on observable equity values is
given by
QT [µ,
2
]=
1
ln(2⇡
2
2
)
T
1 X Rt ( 2 )
2T t=1
(µ
2
29
1
2
2
)
2
n
1X
ln g(St ,
T t=1
2
)
T
1X
ln
T t=1
dt ( 2 )
,
where implicit returns
Rt ( 2 ) = ln(g[St ,
2
])
ln(g[St 1 ,
2
]),
can be obtained using the Black and Scholes formula and a given value of 2 . Estimation of
(µ, 2 ) then proceeds by maximizing QT [µ, 2 ].
Since estimation of µ is not a priority the first-step is often to concentrate out µ, which
yields
T
2
2
1X
µT ( 2 ) =
Rt ( 2 ) +
= R̄T ( 2 ) + ,
T t=1
2
2
and the log-likelihood based on the observable equity values becomes
2
QT [ ] =
1
log(2⇡
2
2
)
T
1 X Rt ( 2 )
2T t=1
R̄T ( 2 )
2
T
1X
log g[St ,
T t=1
2
2
]
T
1X
log
T j=1
dt ( 2 ) .
QT [ 2 ] depends, in several places, on the structural relationship g[St , 2 ], which makes directly
maximizing QT [ 2 ] numerically unstable. As in section Section 5.1, we denote the problematic occurrences of 2 in QT [ 2 ] due to the structural relationship g[St , 2 ] by ⌫( 2 ); note,
⌫( 2 ) = 2 and the di↵erence between the two occurrences of 2 is for notational purposes.
The concentrated log-likelihood function then becomes
2
2
QT [ , ⌫( )] =
1
ln(2⇡
2
T
1X
ln
T t=1
2
T
1 X Rt (⌫( 2 ))
2T t=1
)
R̄T (⌫( 2 ))
2
T
1X
ln g[St , ⌫( 2 )]
T t=1
2
dt (⌫( 2 )) .
Defining
˜T2 [⌫( 2 )]
T
1X
=
(Rj (⌫( 2 ))
T j=1
an estimator of
2
R̄T (⌫( 2 )))2 and AT [⌫( 2 )] = 2
@QT [ 2 , ⌫( 2 )] @⌫( 2 )
,
@⌫
@ 2
can be obtained as the solution to the log-likelihood first-order conditions
0=
1
2
+
1
4
˜T2 [⌫( 2 )] + AT [⌫( 2 )].
Solving the above equation is equivalent to solving the estimating equation 0 = qT [ 2 , ⌫( 2 )],
where
qT [ 2 , ⌫( 2 )] =
4
AT [⌫( 2 )]
2
+ ˜T2 [⌫( 2 )].
Directly solving 0 = qT [ 2 , ⌫( 2 )] to estimate 2 can be cumbersome, and a popular alternative, due to Kealhofer, Mcquown and Vasicek and dubbed the KMV iterative method, is to
30
base estimation of
2
on
˜T2 [⌫( 2 )] =
T
1X
(Rj (⌫( 2 ))
T j=1
R̄T (⌫( 2 )))2 .
Given a starting value ˆ 2(1) , for k > 1, the KMV iterative method updates its estimates of
by calculating
ˆ 2(k) = ˜T2 [⌫(ˆ 2(k
1)
)] =
T
1X
(Rt (⌫(ˆ 2(k
T t=1
1)
R̄T (⌫(ˆ 2(k
))
1)
2
)))2 ,
and iterating till convergence. This iterative procedure is often much simpler than one based on
solving qT [ 2 , ⌫( 2 )] = 0 since it completely neglects the influence of AT [⌫( 2 )] on the estimates
of 2 . FPR demonstrate that the iterative KMV approach coincides with the latent backfitting
estimator proposed by Pastorello et al. (2003) (hereafter, PPR).
While much simpler than maximum likelihood, the KMV/PPR estimator does not utilize all
of the information in the estimating equation qT [ 2 , ⌫( 2 )] and therefore is not asymptotically
equivalent to the MLE. To this end, FPR use the MBP approach to obtain an estimator that
maintains some of the computational advantages of the KMV/PPR iterative strategy yet still
delivers an estimator that is asymptotically equivalent to the MLE. Given an initial estimator
ˆ 2(1) , at the k-th iteration (k > 1) the MBP estimator solves the following second-order equation
in 2 :
4
AT [⌫(ˆ 2(k
1)
)]
2
+ ˜T2 [⌫(ˆ 2(k
1)
)] = 0.
(42)
An alternative to the KMV/PPR and MBP approaches is the two-step approach discussed
herein. In this context, the two-step approach linearizes the estimating equations qT [ 2 , ⌫( 2 )],
with respect to the cumbersome occurrences of ⌫( 2 ) = 2 , around an initially consistent
estimator. For ˆ 2(1) an initial estimator of 2 , the non-penalized two-step approach estimates
2
by solving
0=
4
AT [⌫(ˆ 2(1) )]
2
+ ˜T2 [⌫(ˆ 2(1) )] + [@AT [⌫(ˆ 2(1) )]/@⌫] 4 (
+ [@ ˜T2 [⌫(ˆ 2(1) )]/@⌫](
2
and, for ↵T a penalty term, the penalized two-step approach estimates
0=
4
AT [⌫(ˆ 2(1) )]
2
2
ˆ 2(1) )
ˆ 2(1) ),
2
by solving
+ ˜T2 [⌫(ˆ 2(1) )] + [@AT [⌫(ˆ 2(1) )]/@⌫] 4 (
2
ˆ 2(1) )
+ [@ ˜T2 [⌫(ˆ 2(1) )]/@⌫](
2
ˆ 2(1) )2 .
2
ˆ 2(1) ) + ↵T (
(43)
(44)
Note that the two-step estimators in equations (43) and (44) require solving a third-order
equation in 2 , whereas the MBP estimator in equation (42) solves a second-order equation.
However, the two-step estimators solve only one third-order equation in 2 , whereas the MBP
estimator requires solving (potentially) many second-order equations in 2 . The computational
merits of both approaches will depend on the quality of the first-step estimator ˆ 2(1) and, in the
31
case of MBP, the strength of the contraction mapping guiding the iterations. For both estimation procedures a convenient starting value can be obtained using the KMV/PPR estimation
procedure.
5.2.2
Simulation Example
To illustrate the usefulness of the efficient two-stage method in the context of the Merton credit
risk model we devise a small Monte Carlo experiment comparing the MBP estimator with
the penalized (respectively, non-penalized) two-step estimator. We construct 1,000 synthetic
samples of 250 and 500 time series observations for daily returns. The firm’s value trajectory is
initialized at 10, 000 and the face value of the firm’s debt is fixed at B = 9, 000. The parameters
are set to µ = .01 and 2 = .09. We focus on estimation of 2 only and so we work directly
with the concentrated log-likelihood function for both estimators.5
The MBP estimator is obtained using a Newton-Raphson approach to solve equation (42).
The penalized (respectively, non-penalized) two-step estimator is obtained using a mix of bisection and interpolation and the penalty term satisfies ↵T / T 1/4 . Both methods use starting
values obtained from the KMV/PPR method. Across the 1,000 synthetic samples we calculate
the mean, median, root mean squared error (RMSE) and mean absolute error (MAE) for the
MBP estimator and the two-step estimators.
The results of the Monte Carlo experiment are contained in Table 5. Table 5 demonstrates
that the two-step estimators and the MBP estimator have similar finite sample properties,
with the penalized two-step estimator having significantly smaller RMSE and MAE. It is also
important to point out that, as with the copula example in Section 4.2.2, the finite sample
properties of the penalized and non-penalized two-step estimator are very similar.
6
Conclusion
The development of nonlinear dynamic models in financial econometrics has given rise to estimation problems that are often viewed as computationally difficult. This potential computational
burden has led to the development of computationally light estimators whose starting point is
often a simple consistent estimator of some instrumental parameters. This first step estimator
can be used either for targeting the structural parameters (Indirect Inference a la Gourieroux,
Monfort and Renault (1993)) or for simplifying estimating equations for the parameters of interest. More often than not, this simplification comes at the price of some loss in efficiency. Not
only do two-step estimators have, in general, an asymptotic distribution that depends on the
distribution of the first step estimator, but even iterations may not be able to restore efficiency
(see PPR and references therein).
FPR demonstrate that the aforementioned inefficiency is caused by disregarding the information contained in (some of) the awkward occurrences of the parameters in the criterion
function. Popular iterative (or two-step) procedures are devised precisely to allow us to overlook these awkward occurrences, possibly at the cost of efficiency loss. The goal of FPR was
to propose efficient iterative estimation procedures whose computational cost, at each step of
5
The proposed simulation design is similar to that of FPR.
32
the iteration, is no higher than those of popular inefficient inference procedures. This goal
was made possible by the fact that their algorithms iterate on the occurrences of the parameters that researchers would like to overlook. In this way, the informational content of these
occurrences was no longer ignored, at least in the limit of the iterative procedure.
In the present paper, we replace the method of iteration by a partial linearization of the
estimating equations around a first step consistent estimator for the parameters that are difficult
to deal with. With respect to the efficient iterative procedure of FPR, the pros and cons of our
two step procedure are as follows.
On the one hand, our approach is not required to compute a sequence of estimators but only
a second step estimator. Our second step, in general, maintains the computational simplicity
associated with each step of the FPR iterations. Moreover, while consistency of the FPR
iterations may break down when their so-called Information Dominance condition is not fulfilled,
our approach does not require such a condition.
On the other hand, linearization, when it is only partial, may be a risky exercise because
it may deliver a solution for the non-linearized portion that is biased, even asymptotically,
by the approximated linearization. Then the consistency property of the estimator may be
lost. In order to hedge against this risk, we develop a strategy of targeting first step consistent
estimators, in the spirit of indirect inference. However, in contrast with indirect inference,
targeting is for us only a complementary tool for enforcing consistency. In particular, we don’t
want the asymptotic variance of our second step estimator to be inefficiently driven by the first
step estimator used for targeting. This is the reason why we must elicit a tuning parameter
(the penalty weight) that goes to infinity, in order to enforce consistency, but not too fast in
order to avoid the efficiency loss that would be produced by contamination of the second step
estimator by the inefficiency of the first step estimator.
Finally, it is worth noting that the strategy developed in this paper may be of more general
interest. While indirect inference has demonstrated the usefulness of targeting instrumental
parameters for simple identification of structural parameters of interest, the recent literature
on multivariate GARCH has stressed that targeting some unconditional moments may be a
safe way to hedge against the risk of numerical instability associated with supposedly efficient
estimators, at least in the presence of high dimensional and/or highly nonlinear optimization
problems. In a companion paper, we demonstrate that for multivariate GARCH models, in
contrast to existing targeting strategies, our penalization/targeting approach can deliver numerically stable estimates with good finite sample properties without the need to sacrifice efficiency. Moreover, as pointed out in our copula example, in addition to unconditional moments,
the relatively simple and robust estimators of the marginal distributions can often provide a
useful target.
References
T. Bollerslev and J. M. Wooldridge. Quasi-maximum likelihood estimation and inference in
dynamic models with time-varying covariances. Econometric Reviews, 11(2):143–172, 1992.
B. Crepon, F. Kramarz, and A. Trognon. Parameters of interest, nuisance parameters and
33
orthogonality conditions an application to autoregressive error component models. Journal
of Econometrics, 82(1):135 – 156, 1997.
J.-C. Duan. Maximum likelihood estimation using price data of the derivative contract. Mathematical Finance, 4(2):155–167, 1994.
J.-C. Duan. Correction: Maximum likelihood estimation using price data of the derivative
contract. Mathematical Finance, 10(4):461–462, 2000.
R. Engle. Dynamic conditional correlation. Journal of Business an Economic Statistics, 20(3):
339–350, 2002.
R. Engle and J. Mezrich. Garch for groups. RISK, 9(8):36 – 40, 1996.
Y. Fan, S. Pastorello, and E. Renault. Maximization by parts in extremum estimation. The
Econometrics Journal, 2015. Forthcoming.
C. Gourieroux, A. Monfort, and E. Renault. Indirect inference. Journal of Applied Econometrics, 8:S85–S118, 1993. Special Issue on Econometric Inference Using Simulation Techniques.
C. Gourieroux, A. Monfort, and E. Renault. Two-stage generalized moment method with
applications to regressions with heteroscedasticity of unknown form. Journal of Statistical
Planning and Inference, 50(1):37 – 6193, 1996. Econometric Methodology, Part III.
L. P. Hansen, J. Heaton, and A. Yaron. Finite-sample properties of some alternative gmm
estimators. Journal of Business an Economic Statistics, 14(3):pp. 262–280, 1996.
H. O. Hartley. The modified gauss-newton method for the fitting of non-linear regression
functions by least squares. Technometrics, 3(2):269–280, 1961.
M. Hatanaka. An efficient two-step estimator for the dynamic adjustment model with autoregressive errors. Journal of Econometrics, 2(3):199–220, 1974.
H. Joe. Multivariate models and dependence concepts, volume 73. London: Chapman an Hall,
1997.
Y. Liu and R. Luger. Efficient estimation of copula-garch models. Computational Statistics
an Data Analysis, 53(6):2284 – 2297, 2009. The Fourth Special Issue on Computational
Econometrics.
R. C. Merton. On the pricing of corporate debt: The risk structure of interest rates. The
Journal of Finance, 29(2):449–470, 1974.
W. K. Newey and D. L. McFadden. Large sample estimation and hypothesis testing. In R. F.
Engle and D. L. McFadden, editors, Handbook of Econometrics, volume 4, pages 2111 – 2245.
1994.
D. Noureldin, N. Shephard, and K. Sheppard. Multivariate rotated ARCH models. Journal of
Econometrics, 179(1):16 – 30, 2014.
34
A. Pagan. Two stage and related estimators and their applications. The Review of Economic
Studies, 53(4):517–538, 1986.
A. Pakes and D. Pollard. Simulation and the asymptotics of optimization estimators. Econometrica, 57(5):pp. 1027–1057, 1989.
J. Pan. The jump-risk premia implicit in options: evidence from an integrated time-series
study. Journal of Financial Economics, 63(1):3 – 50, 2002.
S. Pastorello, V. Patilea, and E. Renault. Iterative and recursive estimation in structural
nonadaptive models [with comments, rejoinder]. Journal of Business an Economic Statistics,
21(4):449–482, 2003.
A. J. Patton. Copula-based models for financial time series. In T. Andersen, R. Davis, J.P. Kreiss, and T. Mikosch, editors, Handbook of Financial Time Series, pages 767 – 785.
Springer Verlag, 2009.
E. Renault and N. Touzi. Option hedging and implied volatilities in a stochastic volatility
model1. Mathematical Finance, 6(3):279–302, 1996.
P. M. Robinson. The stochastic di↵erence between econometric statistics. Econometrica, 56
(3):531–548, 1988.
J. H. Shih and T. A. Louis. Inferences on the association parameter in copula models for
bivariate survival data. Biometrics, 51(4):1384–1399, 1995.
P. X.-K. Song. Multivariate dispersion models generated from gaussian copula. Scandinavian
Journal of Statistics, 27(2):305–320, 2000.
P. X.-K. Song, Y. Fan, and J. D. Kalbfleisch. Maximization by parts in likelihood inference [with
comments, rejoinder]. Journal of the American Statistical Association, 100(472):1145–1167,
2005.
A. Trognon and C. Gourieroux. A note on the efficiency of two-step estimation methods. In
Essays in Honor of Edmond Malinvaud, pages 233–248. MIT Press, 1990.
A
Regularity Conditions for Extremum Estimators
In all the applications considered in this paper, the estimating equations fT (✓) = qT [✓, ⌫(✓)] of
interest are obtained as first order conditions of some extremum estimation program:
✓ˆT = arg max QT [✓, ⌫(✓)]
✓2⇥
so that
qT [✓, ⌫(✓)] =
@QT [✓, ⌫(✓)] @⌫ 0 (✓) @QT [✓, ⌫(✓)]
+
@✓
@✓
@⌫
35
(45)
It is worth noting that by contrast with the possibly more general framework mentioned in
the introduction, we have introduced a simplification by considering only a fixed known function
⌫(✓) instead of a more general sample dependent function ⌫T (✓). This may look restrictive since
the function ⌫T (✓) may typically show up when profiling out some specific occurrences of some
components of ✓ and thus be computed as data dependent. However, it must be kept in mind
that the di↵erence is more notational than real since a general objective function Q⇤T [✓, ⌫T (✓)]
may always be rewritten QT [✓, ✓] with a new function defined from Q⇤T [., .] and ⌫T (.) by:
QT [✓, ✓⇤ ] = Q⇤T [✓, ⌫T (✓⇤ )]
(46)
This remark actually shows that we could always choose ⌫(✓) = ✓. We prefer to keep the
notation ⌫(✓) for the sake of notational transparency. In most cases, ⌫(✓) will be nothing but
a sub-vector of ✓. However, while we will keep in mind that (45) is actually not less general
than (46), we will make explicit how the regularity conditions must be interpreted when ⌫T (✓)
is actually a sample-dependent consistent estimator of some underlying unknown true ⌫ 0 (✓).
In the simple set up of (45),the maintained regularity conditions are the following.
R1. The following are satisfied:
(1) ⇥ ⇢ Rp and
⇢ Rq are two compact parameters spaces.
(2) ⌫(.) is a continuous function from ⇥ to , twice continuously di↵erentiable on the interior
of ⇥.
(3) ✓0 2 Int(⇥), interior set of ⇥, and ⌫ 0 = ⌫(✓0 ) 2 Int( ), interior set of .
R2. QT [✓, ⌫] converges in probability towards a nonstochastic function Q1 [✓, ⌫] uniformly on
(✓, ⌫) 2 ⇥ ⇥ .
R3. The function ✓ 7 ! Q1 [✓, ⌫(✓)] attains a unique global maximum on ⇥ at ✓ = ✓0 , unique
solution of the equations q1 [✓, ⌫(✓)] = 0, where
q1 [✓, ⌫(✓)] =
@Q1 [✓, ⌫(✓)] @⌫ 0 (✓) @Q1 [✓, ⌫(✓)]
+
@✓
@✓
@⌫
˚ ˚.
R4. The function QT [✓, ⌫] is twice continuously di↵erentiable on ⇥⇥
R5. The following are satisfied
@ 2 QT ( )
@ @ 0
(1) With 0 = (✓0 , ⌫ 0 ), the second derivative
a non stochastic matrix D( ).
2
0
converges uniformly on
˚ ˚ towards
2 ⇥⇥
QT ( )
(2) The matrix D✓✓ ( 0 ) = P limT =1 @ @✓@✓
(where 00 = (✓00 , ⌫ 00 )) is negative definite.
0
p h @QT ( 0 ) @⌫ 0 0 @QT ( 0 ) i
R6.
T
+ @✓ (✓ ). @⌫
converges in distribution towards a normal distribution
@✓
with zero mean and variance ⌦.
36
It is worth reinterpreting these regularity conditions when the objective function QT is
actually deduced from another function Q⇤T as in (46). Note that in this case, ⌫(.) is just
the identity function (⌫(✓) = ✓, ⇥ = ), making trivial all maintained assumptions about
⌫. However, it must be kept in mind that the role of the data dependent function ⌫T (.) will
typically be the consistent estimation of some true unknown function ⌫ 0 (.). Then, the above
regularity conditions can be rewritten identical by only replacing the functions QT [✓, ⌫] and
⌫(.) by the functions Q⇤T [✓, ⌫] and ⌫ 0 (.) . Only the limit arguments involving the function ⌫ 0 (.)
have to be revisited to take into account its consistent estimation. We will basically rewrite
condition R2 and R6 as follows:
R2*. The following are satisfied:
(1) Q⇤T [✓, ⌫] converges in probability towards a non-stochastic function Q⇤1 [✓, ⌫] uniformly on
(✓, ⌫) 2 ⇥ ⇥ .
(2) ⌫T (✓) converges in probability towards ⌫ 0 (✓) uniformly on ✓ 2 ⇥.
i
p h
(✓ 0 ,✓ 0 )
(✓ 0 ,✓ 0 )
R6*. T @QT @✓
+ @QT@✓
, where QT [✓, ✓⇤ ] = Q⇤T [✓, ⌫T (✓⇤ )],converges in distribution
⇤
towards a normal distribution with zero mean and variance ⌦.
Obviously, a more primitive condition for R6* should be basedpof an assumption of joint
asymptotic normality, involving not only the score function but also T (⌫T (✓0 ) ⌫ 0 (✓0 )) whose
impact on the asymptotic distribution would be deduced from a Taylor expansion
p @Q⇤T (✓0 , ⌫T (✓0 )) p @Q⇤T (✓0 , ⌫ 0 (✓0 )) @ 2 Q⇤T (✓0 , ⌫ 0 (✓0 )) p
T
= T
+
. T ⌫T (✓0 )
@
@
@ @⌫ 0
⌫ 0 (✓0 )
While this more specific set up would not introduce any theoretical complication, we omit it
throughout for sake of expositional simplicity.
B
Proofs
B.1
Proof of Theorem 2.1
Part (i) Asymptotic equivalence between ✓ˆT and ✓ˆT⇤ :
The first order conditions that characterize ✓ˆT⇤ can be written:
qT⇤ [✓ˆT⇤ , ⌫˜T ] = 0
with
qT⇤ [✓, ⌫˜T ] =
@QT [✓, ⌫˜T ] @ 2 QT [✓, ⌫˜T ]
+
. [⌫(✓)
@✓
@✓@⌫ 0
⌫˜T ] +
@⌫ 0 (✓) @QT [✓, ⌫˜T ]
.
@✓
@⌫ 0
@⌫ 0 (✓)
JT (✓0 ) [⌫(✓)
@✓
Thus, by comparing with (2):
qT⇤ [✓, ⌫˜T ] = qT [✓, ⌫˜T ] +
@qT [✓, ⌫˜T ]
. [⌫(✓)
@⌫ 0
37
⌫˜T ]
⇠T (✓)
⌫˜T ]
with

@⌫ 0 (✓) @ 2 QT [✓, ⌫˜T ]
⇠T (✓) =
+ JT (✓0 ) [⌫(✓)
@✓
@⌫@⌫ 0
Hence,
i
@qT [✓ˆT⇤ , ⌫˜T ] h ˆ⇤
.
⌫(
✓
)
⌫
˜
0 = qT [✓ˆT⇤ , ⌫˜T ] +
T
T
@⌫ 0
⇣ p ⌘
⇠T (✓ˆT⇤ ) = oP 1/ T
with
⌫˜T ]
⇠T (✓ˆT⇤ )
by virtue of Assumption A3, since ✓ˆT⇤ is root-T consistent. We will see in section 3 that,
whenever consistent, an estimator ˚
✓T solution of
hT (˚
✓T ) = 0
with
hT (✓) = qT [✓, ⌫˜T ] +
@qT [✓, ⌫˜T ]
. [⌫(✓)
@⌫ 0
⌫˜T ]
being asymptotically equivalent to ✓ˆT . Thus, by application of Theorem 3.3 of Pakes and
Pollard (1989), it is also the case for a solution ✓ˆT⇤ of
⇣ p ⌘
hT (✓ˆT⇤ ) = oP 1/ T .
Note that we can apply Theorem
p 3.3 of Pakes and Pollard (1989) in particular because, by
virtue of Assumptions A2 and A3, T hT (✓0 ) is asymptotically normal.
Part (ii): Asymptotic equivalence between ✓ˆT⇤ and ✓ˆText
By definition, ✓ˆText is solution of first order conditions:
gT (✓ˆText ) = 0
such that
⇣ p ⌘
gT (✓ˆT⇤ ) = oP 1/ T ,
since gT (✓ˆT⇤ ) is a p-dimensional vector whose component j = 1, ..., p is
h
i0
h
i
⌫(✓ˆT⇤ ) ⌫˜T JjT (✓ˆT⇤ ) ⌫(✓ˆT⇤ ) ⌫˜T .
where Jj;T (✓) stands for the matrix of partial derivatives with respect to ✓j of all the coefficients
of the matrix JT (✓). Then, the announced asymptotic equivalence follows again by application
of Theorem 3.3 of Pakes and Pollard (1989).
B.2
Proof of Proposition 3.1
Step 1: We show that ✓˜TP is a consistent estimator of ✓0 .
38
By definition:
0 = qT [✓˜TP , ⌫(✓˜T )] +
@qT ˜P
@⌫
[✓T , ⌫(✓˜T )] 0 (✓˜T )(✓˜TP
0
@⌫
@✓
✓˜T ) + ↵T ✓˜TP
✓˜T
2
ep
(47)
Since the parameter space is compact, we only have to show that for any subsequence of ✓˜TP
¯ we necessarily have ✓¯ = ✓0 . By the
that converges in probability towards some limit value ✓,
regularity conditions (continuity and uniform convergence) we deduce from (47) that
¯ ⌫(✓0 )] + @q1 [✓,
¯ ⌫(✓0 )] @⌫ (✓0 )(✓¯
0 = q1 [✓,
0
@⌫
@✓0
✓0 ) + P lim ↵T ✓˜TP
✓˜T
T =1
2
ep .
(48)
Since PlimT =1 ✓˜T = ✓0 and limT =1 ↵T = 1, (48) implies that PlimT =1 ✓˜TP = ✓0 .
Step 2: We show that
✓ˆT
✓˜TP = OP
⇣
fT (✓ˆT )
h̃PT (✓ˆT )
⌘
= OP
⇣
h̃PT (✓ˆT )
⌘
This result is a direct consequence of Robinson (1988) Theorem 1 if we can show that the
function h̃PT (✓) is conformable to Robinson’s Assumption A2. We have


@ h̃PT (✓)
@qT [✓, ⌫(✓˜T )]
@ @qT
@⌫ ˜
˜
=
+ 0
[✓, ⌫(✓T )]
(✓T )(✓
0
0
0
@✓
@✓
@✓ @⌫
@✓0
⇣
⌘0
@qT
@⌫
+ 0 [✓, ⌫(✓˜T )] 0 (✓˜T ) + 2↵T ✓ ✓˜T
@⌫
@✓
✓˜T ) ⌦ Idp
where, for a (p ⇥ q) matrix A whose coefficients are functions of ✓, we define @A/@✓0 as the
(p ⇥ qp) matrix
⇥ @A1 @A2
q ⇤
... @A
@✓ 0
@✓ 0
@✓ 0
where A1 , A2 , ..., Aq stands for the q columns of the matrix A. Since , by assumption, ✓˜T
oP (1/↵T ), we deduce that, under regularity conditions
P lim
✓0 =
@hT (✓0 )
@q1 [✓0 , ⌫(✓0 )] @q1 0
@⌫
=
+
[✓ , ⌫(✓0 )] 0 (✓0 ) = F,
0
0
0
@✓
@✓
@⌫
@✓
that is by assumption a non-singular matrix. Therefore, we get Assumption A2 of Robinson
(1988) under standard regularity conditions.
Step 3: We show that
✓ˆT
✓˜TP = OP
✓
↵T
39
✓ˆT
✓˜T
2
◆
.
We have
fT (✓ˆT ) = qT [✓ˆT , ⌫(✓ˆT )]
@qT ˆ
@⌫
= qT [✓ˆT , ⌫(✓˜T )] +
[✓T , ⌫(✓˜T )] 0 (✓˜T )(✓ˆT ✓˜T ) + OP
0
@⌫
@✓
✓
◆
2
2
= h̃PT (✓ˆT ) + OP
✓ˆT ✓˜T
↵T ✓ˆT ✓˜T ep .
Therefore,
fT (✓ˆT )
h̃PT (✓ˆT )
= OP
✓
↵T
✓ˆT
✓˜T
which gives the announced result by using the result of Step 2.
B.3
2
◆
✓
✓ˆT
✓˜T
2
◆
,
Proof Theorem 3.1
We just show that the proof of Proposition 3.1. will go through with very minor changes. The
proof of consistency (Step 1) is the same except that equation (48) must now be replaced by
¯ ⌫(✓0 )] +
0 = q1 [✓,
@q1 0
@⌫
[✓ , ⌫(✓0 )] 0 (✓0 )(✓¯
0
@⌫
@✓
✓˜T
✓0 ) + PlimT =1 ↵T ✓T⇤⇤
2
ep
(49)
Obviously, the same consistency argument is a fortiori still valid. Since PlimT =1 ✓˜T = ✓0 and
(1)
limT =1 ↵T = 1, (49) implies that PlimT =1 ✓T = ✓0 . With this new way to partially linearize,
the Jacobian of the estimating equation is simplified as follows
(1)
⇣
@qT [✓, ⌫(✓˜T )] @qT ˜
@hT (✓)
˜T )] @⌫ (✓˜T ) + 2↵T ✓
=
+
[
✓
,
⌫(
✓
T
@✓0
@✓0
@⌫ 0
@✓0
✓˜T
Thus, we still have
⌘0
(1)
lim
@hT (✓0 )
@⌫
@q1 [✓0 , ⌫(✓0 )] @q1 0
=
+
[✓ , ⌫(✓0 )] 0 (✓0 ) = F
0
0
0
@✓
@✓
@⌫
@✓
and thus we can prove a Step 2 exactly as in Proposition 3.1. This Step 2 will tell us that
⇣
⌘
✓ˆT ✓T⇤⇤ = OP fT (✓ˆT ) h⇤T (✓ˆT ) .
We already know from Proposition 3.1 that
fT (✓ˆT )
h̃PT (✓ˆT )
= OP
✓
↵T
✓ˆT
✓˜T
2
◆
Thus, the triangle inequality will give the result if we can also show that
✓
◆
2
(1) ˆ
P ˆ
ˆ
˜
h̃T (✓T ) hT (✓T ) = OP ↵T ✓T ✓T
40
We have
h̃PT (✓ˆT )
(1)
hT (✓ˆT )
=

@qT ˆ
[✓T , ⌫(✓˜T )]
@⌫ 0
@qT ˜
@⌫ ˜ ˆ
[✓T , ⌫(✓˜T )]
(✓T )(✓T
0
@⌫
@✓0
✓˜T )
Assuming that the initial estimating equations qT [✓, ⌫] are twice continuously di↵erentiable
on the interior of the compact set ⇥ ⇥ (see regularity conditions in appendix), we know that:
@qT ˆ
[✓T , ⌫(✓˜T )]
@⌫ 0
⇣
@qT ˜
˜T )] = OP ✓ˆT
[
✓
,
⌫(
✓
T
@⌫ 0
Therefore
h̃PT (✓ˆT )
(1)
hT (✓ˆT )
= OP
since ↵T goes to infinity.
B.4
✓
✓ˆT
✓˜T
2
◆
= OP
✓
↵T
✓˜T
✓ˆT
⌘
✓˜T
2
◆
Proof of Theorem 3.2
We just show that the proof of Proposition 3.1. will go through with some suitable changes.
The proof of consistency (Step 1) is the same except that equation (48) must now be
replaced by:
¯ ⌫(✓0 )] + @q1 [✓,
¯ ⌫(✓0 )](⌫(✓)
¯
0 = q1 [✓,
@⌫ 0
(2)
⌫(✓0 ) + P lim ↵T ⌫(✓T )
T =1
⌫˜T
2
ep
(50)
Obviously, the same kind of consistency argument is still valid. Since PlimT =1 ⌫˜T = ⌫(✓0 ) and
(2)
¯ = ⌫(✓0 ). Therefore we must have
limT =1 ↵T = 1, (50) implies that PlimT =1 ⌫(✓T ) = ⌫(✓)
¯ ⌫(✓0 )] = q1 [✓,
¯ ⌫(✓)]
¯
0 = q1 [✓,
from which we deduce ✓¯ = ✓0 by virtue of Assumption B2.
To get Step 2, we now compute the Jacobian of the estimating equations

(2)
@hT (✓)
@qT [✓, ⌫˜T ]
@ @qT
=
+
[✓, ⌫˜T ] [(⌫(✓) ⌫˜T ) ⌦ Idp ]
@✓0
@✓0
@✓0 @⌫ 0

@qT
@⌫
@⌫
+ 0 [✓, ⌫˜T ] 0 (✓) + 2↵T [⌫(✓) ⌫˜T ]0 0 (✓)ep ep
@⌫
@✓
@✓
Thus, we still have
(2)
lim
@hT (✓0 )
@q1 [✓0 , ⌫(✓0 )] @q1 0
@⌫
=
+
[✓ , ⌫(✓0 )] 0 (✓0 ) = F
0
0
0
@✓
@✓
@⌫
@✓
and thus we can prove a Step 2 exactly as in Proposition 3.1. This Step 2 will tell us that
⇣
⌘
(2)
(2)
✓ˆT ✓T = OP fT (✓ˆT ) hT (✓ˆT )
41
To get the announced result, we now (Step 3) need to show that
✓
(2) ˆ
ˆ
fT (✓T ) hT (✓T ) = OP ↵T ⌫(✓ˆT ) ⌫˜T
2
◆
We have
fT (✓ˆT ) = qT [✓ˆT , ⌫(✓ˆT )]
@qT ˆ
= qT [✓ˆT , ⌫˜T ] +
[✓T , ⌫˜T ].(⌫(✓ˆT )
@⌫ 0
✓
◆
2
(2) ˆ
ˆ
= h T ( ✓T ) + OP
⌫(✓T ) ⌫˜T
⌫˜T ) + OP
↵T ⌫(✓ˆT )
✓
⌫(✓ˆT )
⌫˜T
2
⌫˜T
2
◆
ep
which gives the announced result.
C
Tables
The following section details the Monte Carlo results from the simulation experiments in Section 4.2.2 and Section 5.2.1. In the tables below, MBP stands for the maximization by parts
estimator, P-TS1 is the fully penalized two-step estimator, the two-step estimator with partial
penalization is P-TS2 , the two-stage estimator without penalization is TS1 and the simplified
two-step estimator is TS2 .
42
43
↵1
↵2
⇢
↵1
↵2
⇢
↵1
↵2
⇢
↵1
↵2
⇢
↵1
↵2
⇢
MBP
P-T S1
P-T S2
T S1
T S2
10.0389
100.2121
75.3633
10.03671
100.1963
75.3633
10.0387
100.2106
74.6152
10.03615
100.1919
75.3727
10.0366
100.1881
75.3606
1.0375
103.4928
430.1764
1.0369
103.4614
398.9833
1.0375
103.4881
430.1701
1.0367
103.4473
398.5865
1.0429
103.3290
399.0850
81.2189 9.9658
801.4590 99.6928
2038.4899 74.9160
81.2033 9.9636
801.4120 99.6764
1963.6687 74.9160
81.2172 9.9657
801.4460 99.6922
2038.4766 74.2100
81.1981 9.9634
801.3737 99.6749
1962.7233 74.9204
81.3085 9.9641
801.1743 99.6738
1963.9320 74.9131
0.5655
54.0369
439.7444
0.5650
54.0001
410.1344
0.5655
54.0372
439.7422
0.56507
54.0004
409.9491
0.5694
54.0119
410.2502
10.0047
99.6255
75.0674
59.8332 10.0071
605.4414 99.6437
2008.6820 75.0646
59.6107 10.0048
605.3864 99.6264
2008.3948 75.0646
59.6264 10.0071
605.6657 99.6434
2078.9956 74.3475
59.6130
605.3785
2007.9506
59.8332 10.0058
605.4414 99.6216
2008.6820 75.0615
0.3413
34.5841
431.9198
0.34109
34.5723
402.2031
0.3413
34.5842
431.9187
0.34109
34.5726
402.0887
0.3432
34.5963
402.3241
46.2821
469.1803
2065.2523
46.2837
469.1192
1993.5322
46.2821
469.1822
2065.2499
46.2839
469.1240
1993.2528
46.3782
469.3129
1993.8420
Table 1: µ100 is the estimated Mean, times 100, for the di↵erent estimator, MAE100 is the Monte Carlo mean squared
error, times 1000, and MAE100 is the Monte Carlo mean absolute error, times 1000. The penalization parameter was taken
proportional to T 1/4 . The parameters ↵1 , ↵2 were set equal to .1 and 1 across all sample sizes. The table below fixed ⇢ = .75.
⇢ = .75
µ100
MSE100
MAE100
µ200
MSE200
MAE200
µ300
MSE300
MAE300
44
↵1
↵2
⇢
↵1
↵2
⇢
↵1
↵2
⇢
↵1
↵2
⇢
↵1
↵2
⇢
MBP
P-T S1
P-T S2
T S1
T S2
10.0381
100.3489
95.0797
10.0361
100.3309
95.0797
10.0381
100.3489
94.8122
10.0361
100.3307
95.0799
10.0368
100.3326
95.0740
1.0382
101.3217
0.8194
1.0380
101.3135
0.5909
1.0382
101.3216
0.8194
1.03802
101.3133
0.5908
1.0436
101.6088
0.5900
9.9679
99.5895
94.9604
81.2629
802.9575
72.4880
81.2689
802.8293
61.3679
81.2629
802.9570
72.4880
9.9645
99.6126
94.9894
9.9624
99.5939
94.9894
9.9645
99.6126
94.7308
81.2689 9.9624
802.8277 99.5938
61.3661 94.9895
81.2541
803.7984
61.2780
0.5656
56.7682
0.4688
0.5655
56.7602
0.2910
0.5656
56.7682
0.4688
0.5655
56.7603
0.2909
0.5843
57.6078
0.2925
10.0062
99.8697
94.7739
10.0041
99.8505
95.0280
60.4423 10.0062
608.1178 99.8697
43.7780 95.0280
59.6428 10.0041
604.8638 99.8505
43.5963 95.0280
59.6420
604.96135
54.3549
59.6429
604.8641
43.5954
60.4423 10.0609
608.1178 99.8313
43.7780 94.5256
46.3441
468.8434
46.7765
46.3435
468.8306
37.6353
0.34205
35.0885
0.3468
46.3440
468.8434
46.7766
0.3419 46.3434
35.0822 468.8304
0.2124 37.6357
0.3420
35.0885
0.3468
0.3419
35.0822
0.2124
0.5784 61.1738
59.6184 615.5324
0.5025 56.1719
Table 2: µ100 is the estimated Mean, times 100, for the di↵erent estimator, MAE100 is the Monte Carlo mean squared
error, times 1000, and MAE100 is the Monte Carlo mean absolute error, times 1000. The penalization parameter was taken
proportional to T 1/4 . The parameters ↵1 , ↵2 were set equal to .1 and 1 across all sample sizes. The table below fixed ⇢ = .95.
⇢ = .95
µ100
MSE100 MAE100
µ200 MSE200
MAE200
µ300 MSE300 MAE300
45
↵1
↵2
⇢
↵1
↵2
⇢
↵1
↵2
⇢
↵1
↵2
⇢
↵1
↵2
⇢
MBP
P-T S1
P-T S2
T S1
T S2
10.0367
100.3582
98.5239
10.0357
100.3488
99.8524
10.0367
100.3582
98.4358
10.0357
100.3488
98.5239
10.0372
100.8034
98.0858
1.0393
102.2073
0.0801
1.0392
102.2061
0.0541
1.0393
102.2073
0.0801
1.0392
102.2060
0.0541
1.2314
123.5412
0.2934
81.2839
804.7325
22.7917
81.2881
804.7207
18.6253
81.2839
804.7324
22.7917
81.2881
804.7207
18.6253
88.7035
879.2445
43.3899
9.9631
99.6056
98.4977
9.96217
99.5960
98.4977
9.9631
99.6055
98.4122
9.9621
99.5960
98.4977
10.1446
100.1246
96.4156
112.0659
1108.4347
208.4360
0.5662
57.0553
0.0457
0.5662
57.0553
0.0263
0.5662
57.0552
0.0457
10.0039
99.9360
98.5098
59.6806
600.3720
16.9301
59.6819
600.34731
13.1029
10.0049
99.9456
98.5098
10.0039
99.9361
98.5098
0.3421
34.8572
0.0331
0.3421
34.8534
0.0193
0.3421
34.8571
0.0330
0.3421
34.8534
0.0193
10.2592
2.4063
01.5723 242.56481
94.8641
13.5292
59.6806 10.0049
600.3720 99.9456
16.9308 98.42705
0.5662
59.6819
57.0553 600.34733
0.0263
13.1028
1.7055
165.4135
4.6093
46.3561
468.5121
14.5022
46.3557
468.471
11.3652
46.3560
468.5121
14.5021
46.3557
468.4711
11.3651
143.4284
1439.3493
363.5840
Table 3: µ100 is the estimated Mean, times 100, for the di↵erent estimator, MAE100 is the Monte Carlo mean squared
error, times 1000, and MAE100 is the Monte Carlo mean absolute error, times 1000. The penalization parameter was taken
proportional to T 1/4 . The parameters ↵1 , ↵2 were set equal to .1 and 1 across all sample sizes. The table below fixed ⇢ = .985.
⇢ = .985
µ100
MSE100 MAE100
µ200
MSE200
MAE200
µ300
MSE300
MAE300
Table 4: Relative computing time, in seconds and number of iterations for MBP (in brackets).
⇢ = .75
⇢ = .95
⇢ = .985
T=100
MBP
P-TS
TS
0.0119 [3]
0.0203
0.0207
0.0260
0.0166
0.0164
[6]
0.0922 [43]
0.0137
0.0142
T=200
MBP
P-TS
TS
0.0152 [4]
0.0193
0.0199
0.0450
0.0141
0.0167
[16]
0.1005
0.0138
0.0148
[44]
T=300
MBP
P-TS
TS
0.0169 [4]
0.0184
0.0187
0.1001
0.0143
0.0160
[ 42]
0.1018
0.0145
0.0149
[45]
Table 5: Results for two-step (TS1 , penalized two-step (P-TS1 ) and MBP estimators in the
Merton credit Risk model. MAE is the median absolute error across the simulations multiplied
by 100, and RMSE is the root mean squared error across the simulation multiplied by 100.
TS
T
T=250
T=500
Parameter
= 0.09
= 0.09
Median
0.0889
0.0897
Mean.
0.0888
0.0895
MAE
6.1099
4.7301
RMSE
7.7331
5.8404
Parameter
= 0.09
= 0.09
Median
0.0895
0.0898
Mean.
0.0890
0.0895
MAE
5.9292
4.6715
RMSE
7.5341
5.6284
Parameter
= 0.09
= 0.09
Median
0.0892
0.0894
Mean
0.0888
0.0898
MAE
8.1746
6.7727
RMSE
9.8406
6.6129
P-TS
T
T=250
T=500
MBP
T
T=250
T=500
46
Download